[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-oumi-ai--oumi":3,"tool-oumi-ai--oumi":61},[4,18,26,36,44,53],{"id":5,"name":6,"github_repo":7,"description_zh":8,"stars":9,"difficulty_score":10,"last_commit_at":11,"category_tags":12,"status":17},4358,"openclaw","openclaw\u002Fopenclaw","OpenClaw 是一款专为个人打造的本地化 AI 助手，旨在让你在自己的设备上拥有完全可控的智能伙伴。它打破了传统 AI 助手局限于特定网页或应用的束缚，能够直接接入你日常使用的各类通讯渠道，包括微信、WhatsApp、Telegram、Discord、iMessage 等数十种平台。无论你在哪个聊天软件中发送消息，OpenClaw 都能即时响应，甚至支持在 macOS、iOS 和 Android 设备上进行语音交互，并提供实时的画布渲染功能供你操控。\n\n这款工具主要解决了用户对数据隐私、响应速度以及“始终在线”体验的需求。通过将 AI 部署在本地，用户无需依赖云端服务即可享受快速、私密的智能辅助，真正实现了“你的数据，你做主”。其独特的技术亮点在于强大的网关架构，将控制平面与核心助手分离，确保跨平台通信的流畅性与扩展性。\n\nOpenClaw 非常适合希望构建个性化工作流的技术爱好者、开发者，以及注重隐私保护且不愿被单一生态绑定的普通用户。只要具备基础的终端操作能力（支持 macOS、Linux 及 Windows WSL2），即可通过简单的命令行引导完成部署。如果你渴望拥有一个懂你",349277,3,"2026-04-06T06:32:30",[13,14,15,16],"Agent","开发框架","图像","数据工具","ready",{"id":19,"name":20,"github_repo":21,"description_zh":22,"stars":23,"difficulty_score":10,"last_commit_at":24,"category_tags":25,"status":17},3808,"stable-diffusion-webui","AUTOMATIC1111\u002Fstable-diffusion-webui","stable-diffusion-webui 是一个基于 Gradio 构建的网页版操作界面，旨在让用户能够轻松地在本地运行和使用强大的 Stable Diffusion 图像生成模型。它解决了原始模型依赖命令行、操作门槛高且功能分散的痛点，将复杂的 AI 绘图流程整合进一个直观易用的图形化平台。\n\n无论是希望快速上手的普通创作者、需要精细控制画面细节的设计师，还是想要深入探索模型潜力的开发者与研究人员，都能从中获益。其核心亮点在于极高的功能丰富度：不仅支持文生图、图生图、局部重绘（Inpainting）和外绘（Outpainting）等基础模式，还独创了注意力机制调整、提示词矩阵、负向提示词以及“高清修复”等高级功能。此外，它内置了 GFPGAN 和 CodeFormer 等人脸修复工具，支持多种神经网络放大算法，并允许用户通过插件系统无限扩展能力。即使是显存有限的设备，stable-diffusion-webui 也提供了相应的优化选项，让高质量的 AI 艺术创作变得触手可及。",162132,"2026-04-05T11:01:52",[14,15,13],{"id":27,"name":28,"github_repo":29,"description_zh":30,"stars":31,"difficulty_score":32,"last_commit_at":33,"category_tags":34,"status":17},1381,"everything-claude-code","affaan-m\u002Feverything-claude-code","everything-claude-code 是一套专为 AI 编程助手（如 Claude Code、Codex、Cursor 等）打造的高性能优化系统。它不仅仅是一组配置文件，而是一个经过长期实战打磨的完整框架，旨在解决 AI 代理在实际开发中面临的效率低下、记忆丢失、安全隐患及缺乏持续学习能力等核心痛点。\n\n通过引入技能模块化、直觉增强、记忆持久化机制以及内置的安全扫描功能，everything-claude-code 能显著提升 AI 在复杂任务中的表现，帮助开发者构建更稳定、更智能的生产级 AI 代理。其独特的“研究优先”开发理念和针对 Token 消耗的优化策略，使得模型响应更快、成本更低，同时有效防御潜在的攻击向量。\n\n这套工具特别适合软件开发者、AI 研究人员以及希望深度定制 AI 工作流的技术团队使用。无论您是在构建大型代码库，还是需要 AI 协助进行安全审计与自动化测试，everything-claude-code 都能提供强大的底层支持。作为一个曾荣获 Anthropic 黑客大奖的开源项目，它融合了多语言支持与丰富的实战钩子（hooks），让 AI 真正成长为懂上",142651,2,"2026-04-06T23:34:12",[14,13,35],"语言模型",{"id":37,"name":38,"github_repo":39,"description_zh":40,"stars":41,"difficulty_score":32,"last_commit_at":42,"category_tags":43,"status":17},2271,"ComfyUI","Comfy-Org\u002FComfyUI","ComfyUI 是一款功能强大且高度模块化的视觉 AI 引擎，专为设计和执行复杂的 Stable Diffusion 图像生成流程而打造。它摒弃了传统的代码编写模式，采用直观的节点式流程图界面，让用户通过连接不同的功能模块即可构建个性化的生成管线。\n\n这一设计巧妙解决了高级 AI 绘图工作流配置复杂、灵活性不足的痛点。用户无需具备编程背景，也能自由组合模型、调整参数并实时预览效果，轻松实现从基础文生图到多步骤高清修复等各类复杂任务。ComfyUI 拥有极佳的兼容性，不仅支持 Windows、macOS 和 Linux 全平台，还广泛适配 NVIDIA、AMD、Intel 及苹果 Silicon 等多种硬件架构，并率先支持 SDXL、Flux、SD3 等前沿模型。\n\n无论是希望深入探索算法潜力的研究人员和开发者，还是追求极致创作自由度的设计师与资深 AI 绘画爱好者，ComfyUI 都能提供强大的支持。其独特的模块化架构允许社区不断扩展新功能，使其成为当前最灵活、生态最丰富的开源扩散模型工具之一，帮助用户将创意高效转化为现实。",107888,"2026-04-06T11:32:50",[14,15,13],{"id":45,"name":46,"github_repo":47,"description_zh":48,"stars":49,"difficulty_score":32,"last_commit_at":50,"category_tags":51,"status":17},4721,"markitdown","microsoft\u002Fmarkitdown","MarkItDown 是一款由微软 AutoGen 团队打造的轻量级 Python 工具，专为将各类文件高效转换为 Markdown 格式而设计。它支持 PDF、Word、Excel、PPT、图片（含 OCR）、音频（含语音转录）、HTML 乃至 YouTube 链接等多种格式的解析，能够精准提取文档中的标题、列表、表格和链接等关键结构信息。\n\n在人工智能应用日益普及的今天，大语言模型（LLM）虽擅长处理文本，却难以直接读取复杂的二进制办公文档。MarkItDown 恰好解决了这一痛点，它将非结构化或半结构化的文件转化为模型“原生理解”且 Token 效率极高的 Markdown 格式，成为连接本地文件与 AI 分析 pipeline 的理想桥梁。此外，它还提供了 MCP（模型上下文协议）服务器，可无缝集成到 Claude Desktop 等 LLM 应用中。\n\n这款工具特别适合开发者、数据科学家及 AI 研究人员使用，尤其是那些需要构建文档检索增强生成（RAG）系统、进行批量文本分析或希望让 AI 助手直接“阅读”本地文件的用户。虽然生成的内容也具备一定可读性，但其核心优势在于为机器",93400,"2026-04-06T19:52:38",[52,14],"插件",{"id":54,"name":55,"github_repo":56,"description_zh":57,"stars":58,"difficulty_score":10,"last_commit_at":59,"category_tags":60,"status":17},4487,"LLMs-from-scratch","rasbt\u002FLLMs-from-scratch","LLMs-from-scratch 是一个基于 PyTorch 的开源教育项目，旨在引导用户从零开始一步步构建一个类似 ChatGPT 的大型语言模型（LLM）。它不仅是同名技术著作的官方代码库，更提供了一套完整的实践方案，涵盖模型开发、预训练及微调的全过程。\n\n该项目主要解决了大模型领域“黑盒化”的学习痛点。许多开发者虽能调用现成模型，却难以深入理解其内部架构与训练机制。通过亲手编写每一行核心代码，用户能够透彻掌握 Transformer 架构、注意力机制等关键原理，从而真正理解大模型是如何“思考”的。此外，项目还包含了加载大型预训练权重进行微调的代码，帮助用户将理论知识延伸至实际应用。\n\nLLMs-from-scratch 特别适合希望深入底层原理的 AI 开发者、研究人员以及计算机专业的学生。对于不满足于仅使用 API，而是渴望探究模型构建细节的技术人员而言，这是极佳的学习资源。其独特的技术亮点在于“循序渐进”的教学设计：将复杂的系统工程拆解为清晰的步骤，配合详细的图表与示例，让构建一个虽小但功能完备的大模型变得触手可及。无论你是想夯实理论基础，还是为未来研发更大规模的模型做准备",90106,"2026-04-06T11:19:32",[35,15,13,14],{"id":62,"github_repo":63,"name":64,"description_en":65,"description_zh":66,"ai_summary_zh":67,"readme_en":68,"readme_zh":69,"quickstart_zh":70,"use_case_zh":71,"hero_image_url":72,"owner_login":73,"owner_name":74,"owner_avatar_url":75,"owner_bio":76,"owner_company":77,"owner_location":77,"owner_email":78,"owner_twitter":79,"owner_website":80,"owner_url":81,"languages":82,"stars":107,"forks":108,"last_commit_at":109,"license":110,"difficulty_score":32,"env_os":111,"env_gpu":112,"env_ram":113,"env_deps":114,"category_tags":128,"github_topics":130,"view_count":32,"oss_zip_url":77,"oss_zip_packed_at":77,"status":17,"created_at":143,"updated_at":144,"faqs":145,"releases":146},4880,"oumi-ai\u002Foumi","oumi","Easily fine-tune, evaluate and deploy gpt-oss, Qwen3, DeepSeek-R1, or any open source LLM \u002F VLM!","Oumi 是一个专为构建最先进基础模型打造的全流程开源平台，旨在让大语言模型（LLM）和多模态模型（VLM）的微调、评估与部署变得简单高效。无论是热门的 Qwen3、DeepSeek-R1，还是其他开源模型，用户都能通过 Oumi 轻松上手。\n\n它主要解决了开发者在模型定制过程中面临的环境配置复杂、工具链分散以及部署门槛高等痛点，提供了一条从数据处理到最终上线的端到端解决方案。通过统一的接口和自动化流程，Oumi 大幅降低了技术摩擦，让用户能更专注于模型效果本身。\n\n这款工具非常适合 AI 研究人员、算法工程师以及希望深入定制开源模型的开发者使用。如果你需要快速验证新想法或构建专属的行业模型，Oumi 能提供强有力的支持。\n\n在技术亮点方面，Oumi 不仅兼容最新的 Transformers、TRL 和 vLLM 等核心库，还支持 DeepSpeed 加速训练及多种高级对齐算法（如 KTO、DPO）。近期更新更引入了自动超参数调优、数据合成能力以及对 OpenEnv 强化学习环境的支持，甚至允许用户一键将模型部署到 Fireworks.ai 等云端推理服务，真正实现了“开箱即用”的现代","Oumi 是一个专为构建最先进基础模型打造的全流程开源平台，旨在让大语言模型（LLM）和多模态模型（VLM）的微调、评估与部署变得简单高效。无论是热门的 Qwen3、DeepSeek-R1，还是其他开源模型，用户都能通过 Oumi 轻松上手。\n\n它主要解决了开发者在模型定制过程中面临的环境配置复杂、工具链分散以及部署门槛高等痛点，提供了一条从数据处理到最终上线的端到端解决方案。通过统一的接口和自动化流程，Oumi 大幅降低了技术摩擦，让用户能更专注于模型效果本身。\n\n这款工具非常适合 AI 研究人员、算法工程师以及希望深入定制开源模型的开发者使用。如果你需要快速验证新想法或构建专属的行业模型，Oumi 能提供强有力的支持。\n\n在技术亮点方面，Oumi 不仅兼容最新的 Transformers、TRL 和 vLLM 等核心库，还支持 DeepSpeed 加速训练及多种高级对齐算法（如 KTO、DPO）。近期更新更引入了自动超参数调优、数据合成能力以及对 OpenEnv 强化学习环境的支持，甚至允许用户一键将模型部署到 Fireworks.ai 等云端推理服务，真正实现了“开箱即用”的现代化开发体验。","![Oumi Logo](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Foumi-ai_oumi_readme_0d6331fabdc8.png)\n\n[![Documentation](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FDocumentation-oumi-blue.svg)](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Findex.html)\n[![Blog](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBlog-oumi-blue.svg)](https:\u002F\u002Foumi.ai\u002Fblog)\n[![Twitter](https:\u002F\u002Fimg.shields.io\u002Ftwitter\u002Ffollow\u002FOumi_PBC)](https:\u002F\u002Fx.com\u002FOumi_PBC)\n[![Discord](https:\u002F\u002Fimg.shields.io\u002Fdiscord\u002F1286348126797430814?label=Discord)](https:\u002F\u002Fdiscord.gg\u002Foumi)\n[![PyPI version](https:\u002F\u002Fbadge.fury.io\u002Fpy\u002Foumi.svg)](https:\u002F\u002Fbadge.fury.io\u002Fpy\u002Foumi)\n[![License](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-Apache%202.0-blue.svg)](https:\u002F\u002Fopensource.org\u002Flicenses\u002FApache-2.0)\n[![Tests](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Factions\u002Fworkflows\u002Fpretest.yaml\u002Fbadge.svg?branch=main)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Factions\u002Fworkflows\u002Fpretest.yaml)\n[![GPU Tests](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Factions\u002Fworkflows\u002Fgpu_tests.yaml\u002Fbadge.svg?branch=main)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Factions\u002Fworkflows\u002Fgpu_tests.yaml)\n[![GitHub Repo stars](https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Foumi-ai\u002Foumi)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fstargazers)\n[![Code style: black](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fcode%20style-black-000000.svg)](https:\u002F\u002Fgithub.com\u002Fpsf\u002Fblack)\n[![pre-commit](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fpre--commit-enabled-brightgreen?logo=pre-commit)](https:\u002F\u002Fgithub.com\u002Fpre-commit\u002Fpre-commit)\n[![About](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAbout-oumi-blue.svg)](https:\u002F\u002Foumi.ai)\n\n### Everything you need to build state-of-the-art foundation models, end-to-end\n\n\u003Cp align=\"center\">\n  \u003Ca href=\"https:\u002F\u002Ftrendshift.io\u002Frepositories\u002F12865\">\n    \u003Cimg alt=\"GitHub trending\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Foumi-ai_oumi_readme_4a68feb902da.png\" \u002F>\n  \u003C\u002Fa>\n\u003C\u002Fp>\n\n## 🔥 News\n\n- [2026\u002F03] Upgraded to Transformers v5, TRL v0.30, vLLM v0.19, and veRL v0.7 compatibility\n- [2026\u002F03] [MCP Integration Phase 1](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F2234): package scaffold and dependencies for MCP server support\n- [2026\u002F03] New: `oumi deploy` command for deploying oumi models dedicated inference endpoints on fireworks.ai and parasail\n- [2026\u002F03] Added support for Qwen3.5 model family\n- [2026\u002F03] Inference engines received multiple improvements: list_models api, improved error reporting\n- [2026\u002F02] [Preview of using the Oumi Platform and Lambda to fine-tune and deploy a 4B model for user intent classification](https:\u002F\u002Fyoutu.be\u002F0XpfYRpd_FA)\n- [2026\u002F02] [Lambda and Oumi partner for end-to-end custom model development](https:\u002F\u002Fblog.oumi.ai\u002Fp\u002Flambda-and-oumi-partner-for-end-to)\n- [2025\u002F12] [Oumi v0.6.0 released](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Freleases\u002Ftag\u002Fv0.6.0) with Python 3.13 support, `oumi analyze` CLI command, TRL 0.26+ support, and more\n- [2025\u002F12] [WeMakeDevs AI Agents Assemble Hackathon: Oumi webinar on Finetuning for Text-to-SQL](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=6wPikqRZ7bQ&t=3203s)\n- [2025\u002F12] [Oumi co-sponsors WeMakeDevs AI Agents Assemble Hackathon with over 2000 project submissions](https:\u002F\u002Fwww.wemakedevs.org\u002Fhackathons\u002Fassemblehack25)\n- [2025\u002F11] [Oumi v0.5.0 released](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Freleases\u002Ftag\u002Fv0.5) with advanced data synthesis, hyperparameter tuning automation, support for OpenEnv, and more\n- [2025\u002F11] [Example notebook to perform RLVF fine-tuning with OpenEnv](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi%20-%20OpenEnv%20GRPO%20with%20trl.ipynb), an open source library from the Meta PyTorch team for creating, deploying, and distributing agentic RL environments\n- [2025\u002F10] [Oumi v0.4.1](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Freleases\u002Ftag\u002Fv0.4.1) and [v0.4.2](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Freleases\u002Ftag\u002Fv0.4.2) released] with support for Qwen3-VL and Transformers v4.56, data synthesis documentation and examples, and many bug fixes\n\n\u003Cdetails>\n\u003Csummary>Older updates\u003C\u002Fsummary>\n\n- [2025\u002F09] [Oumi v0.4.0 released](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Freleases\u002Ftag\u002Fv0.4.0) with DeepSpeed support, a Hugging Face Hub cache management tool, KTO\u002FVision DPO trainer support\n- [2025\u002F08] Training and inference support for OpenAI's `gpt-oss-20b` and `gpt-oss-120b`: [recipes here](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Frecipes\u002Fgpt_oss)\n- [2025\u002F08] Aug 14 Webinar - [OpenAI's gpt-oss: Separating the Substance from the Hype](https:\u002F\u002Fyoutu.be\u002Fg1PkAV7fXn0).\n- [2025\u002F08] [Oumi v0.3.0 released](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Freleases\u002Ftag\u002Fv0.3.0) with model quantization (AWQ), an improved LLM-as-a-Judge API, and Adaptive Inference\n- [2025\u002F07] Recipe for [Qwen3 235B](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F235b_a22b_together_infer.yaml)\n- [2025\u002F07] July 24 webinar: [\"Training a State-of-the-art Agent LLM with Oumi + Lambda\"](https:\u002F\u002Fyoutu.be\u002Ff3SU_heBP54)\n- [2025\u002F06] [Oumi v0.2.0 released](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Freleases\u002Ftag\u002Fv0.2.0) with support for GRPO fine-tuning, a plethora of new model support, and much more\n- [2025\u002F06] Announcement of [Data Curation for Vision Language Models (DCVLR) competition](https:\u002F\u002Foumi.ai\u002Fblog\u002Fposts\u002Fannouncing-dcvlr) at NeurIPS2025\n- [2025\u002F06] Recipes for training, inference, and eval with the newly released [Falcon-H1](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Frecipes\u002Ffalcon_h1) and [Falcon-E](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Frecipes\u002Ffalcon_e) models\n- [2025\u002F05] Support and recipes for [InternVL3 1B](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Frecipes\u002Fvision\u002Finternvl3)\n- [2025\u002F04] Added support for training and inference with Llama 4 models: Scout (17B activated, 109B total) and Maverick (17B activated, 400B total) variants, including full fine-tuning, LoRA, and QLoRA configurations\n- [2025\u002F04] Recipes for [Qwen3 model family](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Frecipes\u002Fqwen3)\n- [2025\u002F04] Introducing HallOumi: a State-of-the-Art Claim-Verification Model [(technical overview)](https:\u002F\u002Foumi.ai\u002Fblog\u002Fposts\u002Fintroducing-halloumi)\n- [2025\u002F04] Oumi now supports two new Vision-Language models: [Phi4](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Frecipes\u002Fvision\u002Fphi4) and [Qwen 2.5](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b)\n\n\u003C\u002Fdetails>\n\n## 🔎 About\n\nOumi is a fully open-source platform that streamlines the entire lifecycle of foundation models - from data preparation and training to evaluation and deployment. Whether you're developing on a laptop, launching large scale experiments on a cluster, or deploying models in production, Oumi provides the tools and workflows you need.\n\nWith Oumi, you can:\n\n- 🚀 Train and fine-tune models from 10M to 405B parameters using state-of-the-art techniques (SFT, LoRA, QLoRA, GRPO, and more)\n- 🤖 Work with both text and multimodal models (Llama, DeepSeek, Qwen, Phi, and others)\n- 🔄 Synthesize and curate training data with LLM judges\n- ⚡️ Deploy models efficiently with popular inference engines (vLLM, SGLang)\n- 📊 Evaluate models comprehensively across standard benchmarks\n- 🌎 Run anywhere - from laptops to clusters to clouds (AWS, Azure, GCP, Lambda, and more)\n- 🔌 Integrate with both open models and commercial APIs (OpenAI, Anthropic, Vertex AI, Together, Parasail, ...)\n\nAll with one consistent API, production-grade reliability, and all the flexibility you need for research.\n\nLearn more at [oumi.ai](https:\u002F\u002Foumi.ai\u002Fdocs), or jump right in with the [quickstart guide](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fget_started\u002Fquickstart.html).\n\n## 🚀 Getting Started\n\n| **Notebook** | **Try in Colab** | **Goal** |\n|----------|--------------|-------------|\n| **🎯 Getting Started: A Tour** | \u003Ca target=\"_blank\" href=\"https:\u002F\u002Fcolab.research.google.com\u002Fgithub\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi - A Tour.ipynb\">\u003Cimg src=\"https:\u002F\u002Fcolab.research.google.com\u002Fassets\u002Fcolab-badge.svg\" alt=\"Open In Colab\"\u002F>\u003C\u002Fa> | Quick tour of core features: training, evaluation, inference, and job management |\n| **🔧 Model Finetuning Guide** | \u003Ca target=\"_blank\" href=\"https:\u002F\u002Fcolab.research.google.com\u002Fgithub\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi - Finetuning Tutorial.ipynb\">\u003Cimg src=\"https:\u002F\u002Fcolab.research.google.com\u002Fassets\u002Fcolab-badge.svg\" alt=\"Open In Colab\"\u002F>\u003C\u002Fa> | End-to-end guide to LoRA tuning with data prep, training, and evaluation |\n| **📚 Model Distillation** | \u003Ca target=\"_blank\" href=\"https:\u002F\u002Fcolab.research.google.com\u002Fgithub\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi - Distill a Large Model.ipynb\">\u003Cimg src=\"https:\u002F\u002Fcolab.research.google.com\u002Fassets\u002Fcolab-badge.svg\" alt=\"Open In Colab\"\u002F>\u003C\u002Fa> | Guide to distilling large models into smaller, efficient ones |\n| **📋 Model Evaluation** | \u003Ca target=\"_blank\" href=\"https:\u002F\u002Fcolab.research.google.com\u002Fgithub\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi - Evaluation with Oumi.ipynb\">\u003Cimg src=\"https:\u002F\u002Fcolab.research.google.com\u002Fassets\u002Fcolab-badge.svg\" alt=\"Open In Colab\"\u002F>\u003C\u002Fa> | Comprehensive model evaluation using Oumi's evaluation framework |\n| **☁️ Remote Training** | \u003Ca target=\"_blank\" href=\"https:\u002F\u002Fcolab.research.google.com\u002Fgithub\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi - Running Jobs Remotely.ipynb\">\u003Cimg src=\"https:\u002F\u002Fcolab.research.google.com\u002Fassets\u002Fcolab-badge.svg\" alt=\"Open In Colab\"\u002F>\u003C\u002Fa> | Launch and monitor training jobs on cloud (AWS, Azure, GCP, Lambda, etc.) platforms |\n| **📈 LLM-as-a-Judge** | \u003Ca target=\"_blank\" href=\"https:\u002F\u002Fcolab.research.google.com\u002Fgithub\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi - Simple Judge.ipynb\">\u003Cimg src=\"https:\u002F\u002Fcolab.research.google.com\u002Fassets\u002Fcolab-badge.svg\" alt=\"Open In Colab\"\u002F>\u003C\u002Fa> | Filter and curate training data with built-in judges |\n\n## 🔧 Usage\n\n### Installation\n\nChoose the installation method that works best for you:\n\n\u003Cdetails open>\n\u003Csummary>\u003Cb>Using pip (Recommended)\u003C\u002Fb>\u003C\u002Fsummary>\n\n```bash\n# Basic installation\nuv pip install oumi\n\n# With GPU support\nuv pip install 'oumi[gpu]'\n\n# Latest development version\nuv pip install git+https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi.git\n```\n\nDon't have uv? [Install it](https:\u002F\u002Fdocs.astral.sh\u002Fuv\u002Fgetting-started\u002Finstallation\u002F) or use `pip` instead.\n\n\u003C\u002Fdetails>\n\n\u003Cdetails>\n\u003Csummary>\u003Cb>Using Docker\u003C\u002Fb>\u003C\u002Fsummary>\n\n```bash\n# Pull the latest image\ndocker pull ghcr.io\u002Foumi-ai\u002Foumi:latest\n\n# Run oumi commands\ndocker run --gpus all -it ghcr.io\u002Foumi-ai\u002Foumi:latest oumi --help\n\n# Train with a mounted config\ndocker run --gpus all -v $(pwd):\u002Fworkspace -it ghcr.io\u002Foumi-ai\u002Foumi:latest \\\n    oumi train --config \u002Fworkspace\u002Fmy_config.yaml\n```\n\n\u003C\u002Fdetails>\n\n\u003Cdetails>\n\u003Csummary>\u003Cb>Quick Install Script (Experimental)\u003C\u002Fb>\u003C\u002Fsummary>\n\nTry Oumi without setting up a Python environment. This installs Oumi in an isolated environment:\n\n```bash\ncurl -LsSf https:\u002F\u002Foumi.ai\u002Finstall.sh | bash\n```\n\n\u003C\u002Fdetails>\n\nFor more advanced installation options, see the [installation guide](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fget_started\u002Finstallation.html).\n\n### Oumi CLI\n\nYou can quickly use the `oumi` command to train, evaluate, and infer models using one of the existing [recipes](\u002Fconfigs\u002Frecipes):\n\n```shell\n# Training\noumi train -c configs\u002Frecipes\u002Fsmollm\u002Fsft\u002F135m\u002Fquickstart_train.yaml\n\n# Evaluation\noumi evaluate -c configs\u002Frecipes\u002Fsmollm\u002Fevaluation\u002F135m\u002Fquickstart_eval.yaml\n\n# Inference\noumi infer -c configs\u002Frecipes\u002Fsmollm\u002Finference\u002F135m_infer.yaml --interactive\n```\n\nFor more advanced options, see the [training](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fuser_guides\u002Ftrain\u002Ftrain.html), [evaluation](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fuser_guides\u002Fevaluate\u002Fevaluate.html), [inference](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fuser_guides\u002Finfer\u002Finfer.html), and [llm-as-a-judge](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fuser_guides\u002Fjudge\u002Fjudge.html) guides.\n\n### Running Jobs Remotely\n\nYou can run jobs remotely on cloud platforms (AWS, Azure, GCP, Lambda, etc.) using the `oumi launch` command:\n\n```shell\n# GCP\noumi launch up -c configs\u002Frecipes\u002Fsmollm\u002Fsft\u002F135m\u002Fquickstart_gcp_job.yaml\n\n# AWS\noumi launch up -c configs\u002Frecipes\u002Fsmollm\u002Fsft\u002F135m\u002Fquickstart_gcp_job.yaml --resources.cloud aws\n\n# Azure\noumi launch up -c configs\u002Frecipes\u002Fsmollm\u002Fsft\u002F135m\u002Fquickstart_gcp_job.yaml --resources.cloud azure\n\n# Lambda\noumi launch up -c configs\u002Frecipes\u002Fsmollm\u002Fsft\u002F135m\u002Fquickstart_gcp_job.yaml --resources.cloud lambda\n```\n\n**Note:** Oumi is in \u003Cins>beta\u003C\u002Fins> and under active development. The core features are stable, but some advanced features might change as the platform improves.\n\n## 💻 Why use Oumi?\n\nIf you need a comprehensive platform for training, evaluating, or deploying models, Oumi is a great choice.\n\nHere are some of the key features that make Oumi stand out:\n\n- 🔧 **Zero Boilerplate**: Get started in minutes with ready-to-use recipes for popular models and workflows. No need to write training loops or data pipelines.\n- 🏢 **Enterprise-Grade**: Built and validated by teams training models at scale\n- 🎯 **Research Ready**: Perfect for ML research with easily reproducible experiments, and flexible interfaces for customizing each component.\n- 🌐 **Broad Model Support**: Works with most popular model architectures - from tiny models to the largest ones, text-only to multimodal.\n- 🚀 **SOTA Performance**: Native support for distributed training techniques (FSDP, DeepSpeed, DDP) and optimized inference engines (vLLM, SGLang).\n- 🤝 **Community First**: 100% open source with an active community. No vendor lock-in, no strings attached.\n\n## 📚 Examples &  Recipes\n\nExplore the growing collection of ready-to-use configurations for state-of-the-art models and training workflows:\n\n**Note:** These configurations are not an exhaustive list of what's supported, simply examples to get you started. You can find a more exhaustive list of supported [models](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fresources\u002Fmodels\u002Fsupported_models.html), and datasets ([supervised fine-tuning](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fresources\u002Fdatasets\u002Fsft_datasets.html), [pre-training](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fresources\u002Fdatasets\u002Fpretraining_datasets.html), [preference tuning](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fresources\u002Fdatasets\u002Fpreference_datasets.html), and [vision-language finetuning](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fresources\u002Fdatasets\u002Fvl_sft_datasets.html)) in the oumi documentation.\n\n### Qwen Family\n\n| Model | Example Configurations |\n|-------|------------------------|\n| Qwen3-Next 80B A3B | [LoRA](\u002Fconfigs\u002Frecipes\u002Fqwen3_next\u002Fsft\u002F80b_a3b_lora\u002Ftrain.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fqwen3_next\u002Finference\u002F80b_a3b_infer.yaml) • [Inference (Instruct)](\u002Fconfigs\u002Frecipes\u002Fqwen3_next\u002Finference\u002F80b_a3b_instruct_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fqwen3_next\u002Fevaluation\u002F80b_a3b_eval.yaml) |\n| Qwen3 30B A3B | [LoRA](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fsft\u002F30b_a3b_lora\u002Ftrain.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F30b_a3b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fevaluation\u002F30b_a3b_eval.yaml) |\n| Qwen3 32B | [LoRA](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fsft\u002F32b_lora\u002Ftrain.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F32b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fevaluation\u002F32b_eval.yaml) |\n| Qwen3 14B | [LoRA](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fsft\u002F14b_lora\u002Ftrain.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F14b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fevaluation\u002F14b_eval.yaml) |\n| Qwen3 8B | [FFT](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fsft\u002F8b_full\u002Ftrain.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F8b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fevaluation\u002F8b_eval.yaml) |\n| Qwen3 4B | [FFT](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fsft\u002F4b_full\u002Ftrain.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F4b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fevaluation\u002F4b_eval.yaml) |\n| Qwen3 1.7B | [FFT](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fsft\u002F1.7b_full\u002Ftrain.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F1.7b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fevaluation\u002F1.7b_eval.yaml) |\n| Qwen3 0.6B | [FFT](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fsft\u002F0.6b_full\u002Ftrain.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F0.6b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fevaluation\u002F0.6b_eval.yaml) |\n| QwQ 32B | [FFT](\u002Fconfigs\u002Frecipes\u002Fqwq\u002Fsft\u002Ffull_train.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fqwq\u002Fsft\u002Flora_train.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fqwq\u002Fsft\u002Fqlora_train.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fqwq\u002Finference\u002Finfer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fqwq\u002Fevaluation\u002Feval.yaml) |\n| Qwen2.5-VL 3B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Fsft\u002Ffull\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Fsft\u002Flora\u002Ftrain.yaml)• [Inference (vLLM)](configs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Finference\u002Fvllm_infer.yaml) • [Inference](configs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Finference\u002Finfer.yaml) |\n| Qwen2-VL 2B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Fsft\u002Ffull\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Fsft\u002Flora\u002Ftrain.yaml) • [Inference (vLLM)](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Finference\u002Fvllm_infer.yaml) • [Inference (SGLang)](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Finference\u002Fsglang_infer.yaml) • [Inference](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Finference\u002Finfer.yaml) • [Evaluation](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Fevaluation\u002Feval.yaml) |\n\n### 🐋 DeepSeek R1 Family\n\n| Model | Example Configurations |\n|-------|------------------------|\n| DeepSeek R1 671B | [Inference (Together AI)](configs\u002Frecipes\u002Fdeepseek_r1\u002Finference\u002F671b_together_infer.yaml) |\n| Distilled Llama 8B | [FFT](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_llama_8b\u002Ffull_train.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_llama_8b\u002Flora_train.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_llama_8b\u002Fqlora_train.yaml) • [Inference](configs\u002Frecipes\u002Fdeepseek_r1\u002Finference\u002Fdistill_llama_8b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fevaluation\u002Fdistill_llama_8b\u002Feval.yaml) |\n| Distilled Llama 70B | [FFT](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_llama_70b\u002Ffull_train.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_llama_70b\u002Flora_train.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_llama_70b\u002Fqlora_train.yaml) • [Inference](configs\u002Frecipes\u002Fdeepseek_r1\u002Finference\u002Fdistill_llama_70b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fevaluation\u002Fdistill_llama_70b\u002Feval.yaml) |\n| Distilled Qwen 1.5B | [FFT](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_qwen_1_5b\u002Ffull_train.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_qwen_1_5b\u002Flora_train.yaml) • [Inference](configs\u002Frecipes\u002Fdeepseek_r1\u002Finference\u002Fdistill_qwen_1_5b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fevaluation\u002Fdistill_qwen_1_5b\u002Feval.yaml) |\n| Distilled Qwen 32B | [LoRA](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_qwen_32b\u002Flora_train.yaml) • [Inference](configs\u002Frecipes\u002Fdeepseek_r1\u002Finference\u002Fdistill_qwen_32b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fevaluation\u002Fdistill_qwen_32b\u002Feval.yaml) |\n\n### 🦙 Llama Family\n\n| Model | Example Configurations |\n|-------|------------------------|\n| Llama 4 Scout Instruct 17B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama4\u002Fsft\u002Fscout_instruct_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fllama4\u002Fsft\u002Fscout_instruct_lora\u002Ftrain.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fllama4\u002Fsft\u002Fscout_instruct_qlora\u002Ftrain.yaml) • [Inference (vLLM)](\u002Fconfigs\u002Frecipes\u002Fllama4\u002Finference\u002Fscout_instruct_vllm_infer.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fllama4\u002Finference\u002Fscout_instruct_infer.yaml) • [Inference (Together.ai)](\u002Fconfigs\u002Frecipes\u002Fllama4\u002Finference\u002Fscout_instruct_together_infer.yaml) |\n| Llama 4 Scout 17B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama4\u002Fsft\u002Fscout_base_full\u002Ftrain.yaml)  |\n| Llama 3.1 8B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F8b_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F8b_lora\u002Ftrain.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F8b_qlora\u002Ftrain.yaml) • [Pre-training](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fpretraining\u002F8b\u002Ftrain.yaml) • [Inference (vLLM)](configs\u002Frecipes\u002Fllama3_1\u002Finference\u002F8b_rvllm_infer.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Finference\u002F8b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fevaluation\u002F8b_eval.yaml) |\n| Llama 3.1 70B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F70b_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F70b_lora\u002Ftrain.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F70b_qlora\u002Ftrain.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Finference\u002F70b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fevaluation\u002F70b_eval.yaml) |\n| Llama 3.1 405B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F405b_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F405b_lora\u002Ftrain.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F405b_qlora\u002Ftrain.yaml) |\n| Llama 3.2 1B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fsft\u002F1b_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fsft\u002F1b_lora\u002Ftrain.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fsft\u002F1b_qlora\u002Ftrain.yaml) • [Inference (vLLM)](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Finference\u002F1b_vllm_infer.yaml) • [Inference (SGLang)](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Finference\u002F1b_sglang_infer.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Finference\u002F1b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fevaluation\u002F1b_eval.yaml) |\n| Llama 3.2 3B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fsft\u002F3b_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fsft\u002F3b_lora\u002Ftrain.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fsft\u002F3b_qlora\u002Ftrain.yaml) • [Inference (vLLM)](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Finference\u002F3b_vllm_infer.yaml) • [Inference (SGLang)](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Finference\u002F3b_sglang_infer.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Finference\u002F3b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fevaluation\u002F3b_eval.yaml) |\n| Llama 3.3 70B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama3_3\u002Fsft\u002F70b_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_3\u002Fsft\u002F70b_lora\u002Ftrain.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_3\u002Fsft\u002F70b_qlora\u002Ftrain.yaml) • [Inference (vLLM)](\u002Fconfigs\u002Frecipes\u002Fllama3_3\u002Finference\u002F70b_vllm_infer.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fllama3_3\u002Finference\u002F70b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fllama3_3\u002Fevaluation\u002F70b_eval.yaml) |\n| Llama 3.2 Vision 11B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Fsft\u002F11b_full\u002Ftrain.yaml) • [Inference (vLLM)](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Finference\u002F11b_rvllm_infer.yaml) • [Inference (SGLang)](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Finference\u002F11b_sglang_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Fevaluation\u002F11b_eval.yaml) |\n\n### 🦅 Falcon family\n\n| Model | Example Configurations |\n|-------|------------------------|\n| [Falcon-H1](https:\u002F\u002Fhuggingface.co\u002Fcollections\u002Ftiiuae\u002Ffalcon-h1-6819f2795bc406da60fab8df) | [FFT](\u002Fconfigs\u002Frecipes\u002Ffalcon_h1\u002Fsft\u002F) • [Inference](\u002Fconfigs\u002Frecipes\u002Ffalcon_h1\u002Finference\u002F) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Ffalcon_h1\u002Fevaluation\u002F) |\n| [Falcon-E (BitNet)](https:\u002F\u002Fhuggingface.co\u002Fcollections\u002Ftiiuae\u002Ffalcon-edge-series-6804fd13344d6d8a8fa71130) | [FFT](\u002Fconfigs\u002Frecipes\u002Ffalcon_e\u002Fsft\u002F) • [DPO](\u002Fconfigs\u002Frecipes\u002Ffalcon_e\u002Fdpo\u002F) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Ffalcon_e\u002Fevaluation\u002F) |\n\n### 💎 Gemma 3 Family\n\n| Model | Example Configurations |\n|-------|------------------------|\n| Gemma 3 4B Instruct | [FFT](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Fsft\u002F4b_full\u002Ftrain.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Finference\u002F4b_instruct_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Fevaluation\u002F4b\u002Feval.yaml) |\n| Gemma 3 12B Instruct | [LoRA](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Fsft\u002F12b_lora\u002Ftrain.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Finference\u002F12b_instruct_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Fevaluation\u002F12b\u002Feval.yaml) |\n| Gemma 3 27B Instruct | [LoRA](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Fsft\u002F27b_lora\u002Ftrain.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Finference\u002F27b_instruct_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Fevaluation\u002F27b\u002Feval.yaml) |\n\n### 🦉 OLMo 3 Family\n\n| Model | Example Configurations |\n|-------|------------------------|\n| OLMo 3 7B Instruct | [FFT](\u002Fconfigs\u002Frecipes\u002Folmo3\u002Fsft\u002F7b_full\u002Ftrain.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Folmo3\u002Finference\u002F7b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Folmo3\u002Fevaluation\u002F7b\u002Feval.yaml) |\n| OLMo 3 32B Instruct | [LoRA](\u002Fconfigs\u002Frecipes\u002Folmo3\u002Fsft\u002F32b_lora\u002Ftrain.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Folmo3\u002Finference\u002F32b_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Folmo3\u002Fevaluation\u002F32b\u002Feval.yaml) |\n\n### 🎨 Vision Models\n\n| Model | Example Configurations |\n|-------|------------------------|\n| Llama 3.2 Vision 11B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Fsft\u002F11b_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Fsft\u002F11b_lora\u002Ftrain.yaml) • [Inference (vLLM)](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Finference\u002F11b_rvllm_infer.yaml) • [Inference (SGLang)](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Finference\u002F11b_sglang_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Fevaluation\u002F11b_eval.yaml) |\n| LLaVA 7B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllava_7b\u002Fsft\u002Ftrain.yaml) • [Inference (vLLM)](configs\u002Frecipes\u002Fvision\u002Fllava_7b\u002Finference\u002Fvllm_infer.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllava_7b\u002Finference\u002Finfer.yaml) |\n| Phi3 Vision 4.2B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fphi3\u002Fsft\u002Ffull\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fphi3\u002Fsft\u002Flora\u002Ftrain.yaml) • [Inference (vLLM)](configs\u002Frecipes\u002Fvision\u002Fphi3\u002Finference\u002Fvllm_infer.yaml) |\n| Phi4 Vision 5.6B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fphi4\u002Fsft\u002Ffull\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fphi4\u002Fsft\u002Flora\u002Ftrain.yaml) • [Inference (vLLM)](configs\u002Frecipes\u002Fvision\u002Fphi4\u002Finference\u002Fvllm_infer.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fphi4\u002Finference\u002Finfer.yaml) |\n| Qwen2-VL 2B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Fsft\u002Ffull\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Fsft\u002Flora\u002Ftrain.yaml) • [Inference (vLLM)](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Finference\u002Fvllm_infer.yaml) • [Inference (SGLang)](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Finference\u002Fsglang_infer.yaml) • [Inference](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Finference\u002Finfer.yaml) • [Evaluation](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Fevaluation\u002Feval.yaml) |\n| Qwen3-VL 2B | [Inference](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen3_vl_2b\u002Finference\u002Finfer.yaml) |\n| Qwen3-VL 4B | [Inference](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen3_vl_4b\u002Finference\u002Finfer.yaml) |\n| Qwen3-VL 8B | [Inference](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen3_vl_8b\u002Finference\u002Finfer.yaml) |\n| Qwen2.5-VL 3B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Fsft\u002Ffull\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Fsft\u002Flora\u002Ftrain.yaml)• [Inference (vLLM)](configs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Finference\u002Fvllm_infer.yaml) • [Inference](configs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Finference\u002Finfer.yaml) |\n| SmolVLM-Instruct 2B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fsmolvlm\u002Fsft\u002Ffull\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fsmolvlm\u002Fsft\u002Flora\u002Ftrain.yaml) |\n\n### 🔍 Even more options\n\nThis section lists all the language models that can be used with Oumi. Thanks to the integration with the [🤗 Transformers](https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Ftransformers) library, you can easily use any of these models for training, evaluation, or inference.\n\nModels prefixed with a checkmark (✅) have been thoroughly tested and validated by the Oumi community, with ready-to-use recipes available in the [configs\u002Frecipes](configs\u002Frecipes) directory.\n\n\u003Cdetails>\n\u003Csummary>📋 Click to see more supported models\u003C\u002Fsummary>\n\n#### Instruct Models\n\n| Model | Size | Paper | HF Hub  | License  | Open [^1] |\n|-------|------|-------|---------|----------|------|\n| ✅ SmolLM-Instruct | 135M\u002F360M\u002F1.7B | [Blog](https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fsmollm) | [Hub](https:\u002F\u002Fhuggingface.co\u002FHuggingFaceTB\u002FSmolLM-135M-Instruct) | Apache 2.0 | ✅ |\n| ✅ DeepSeek R1 Family | 1.5B\u002F8B\u002F32B\u002F70B\u002F671B | [Blog](https:\u002F\u002Fapi-docs.deepseek.com\u002Fnews\u002Fnews250120) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fdeepseek-ai\u002FDeepSeek-R1) | MIT | ❌ |\n| ✅ Llama 3.1 Instruct | 8B\u002F70B\u002F405B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-3.1-70b-instruct) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama3\u002Flicense\u002F) | ❌  |\n| ✅ Llama 3.2 Instruct | 1B\u002F3B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-3.2-3b-instruct) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama3\u002Flicense\u002F) | ❌  |\n| ✅ Llama 3.3 Instruct | 70B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-3.3-70b-instruct) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama3\u002Flicense\u002F) | ❌  |\n| ✅ Phi-3.5-Instruct | 4B\u002F14B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.14219) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FPhi-3.5-mini-instruct) | [License](https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FPhi-3-mini-4k-instruct\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| ✅ Qwen3 | 0.6B-32B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2505.09388) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3-32B) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| Qwen2.5-Instruct | 0.5B-70B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.16609) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen2.5-7B-Instruct) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| OLMo 2 Instruct | 7B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.00838) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fallenai\u002FOLMo-2-1124-7B) | Apache 2.0 | ✅ |\n| ✅ OLMo 3 Instruct | 7B\u002F32B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.00838) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fallenai\u002FOLMo-3-7B-Instruct) | Apache 2.0 | ✅ |\n| MPT-Instruct | 7B | [Blog](https:\u002F\u002Fwww.mosaicml.com\u002Fblog\u002Fmpt-7b) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmosaicml\u002Fmpt-7b-instruct) | Apache 2.0 | ✅ |\n| Command R | 35B\u002F104B | [Blog](https:\u002F\u002Fcohere.com\u002Fblog\u002Fcommand-r7b) | [Hub](https:\u002F\u002Fhuggingface.co\u002FCohereForAI\u002Fc4ai-command-r-plus) | [License](https:\u002F\u002Fcohere.com\u002Fc4ai-cc-by-nc-license) | ❌ |\n| Granite-3.1-Instruct | 2B\u002F8B | [Paper](https:\u002F\u002Fgithub.com\u002Fibm-granite\u002Fgranite-3.0-language-models\u002Fblob\u002Fmain\u002Fpaper.pdf) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fibm-granite\u002Fgranite-3.1-8b-instruct) | Apache 2.0 | ❌ |\n| Gemma 2 Instruct | 2B\u002F9B | [Blog](https:\u002F\u002Fai.google.dev\u002Fgemma) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fgoogle\u002Fgemma-2-2b-it) | [License](https:\u002F\u002Fai.google.dev\u002Fgemma\u002Fterms) | ❌ |\n| ✅ Gemma 3 Instruct | 4B\u002F12B\u002F27B | [Blog](https:\u002F\u002Fai.google.dev\u002Fgemma) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fgoogle\u002Fgemma-3-27b-it) | [License](https:\u002F\u002Fai.google.dev\u002Fgemma\u002Fterms) | ❌ |\n| DBRX-Instruct | 130B MoE | [Blog](https:\u002F\u002Fwww.databricks.com\u002Fblog\u002Fintroducing-dbrx-new-state-art-open-llm) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fdatabricks\u002Fdbrx-instruct) | Apache 2.0 | ❌ |\n| Falcon-Instruct | 7B\u002F40B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2306.01116) | [Hub](https:\u002F\u002Fhuggingface.co\u002Ftiiuae\u002Ffalcon-7b-instruct) | Apache 2.0 | ❌  |\n| ✅ Llama 4 Scout Instruct | 17B (Activated) 109B (Total) | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-4-Scout-17B-16E-Instruct) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama4\u002Flicense\u002F) | ❌  |\n| ✅ Llama 4 Maverick Instruct | 17B (Activated) 400B (Total) | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-4-Maverick-17B-128E-Instruct) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama4\u002Flicense\u002F) | ❌  |\n\n#### Vision-Language Models\n\n| Model | Size | Paper | HF Hub | License | Open |\n|-------|------|-------|---------|----------|------|\n| ✅ Llama 3.2 Vision | 11B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-3.2-11b-vision) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama3\u002Flicense\u002F) | ❌  |\n| ✅ LLaVA-1.5 | 7B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.03744) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fllava-hf\u002Fllava-1.5-7b-hf) | [License](https:\u002F\u002Fai.meta.com\u002Fllama\u002Flicense) | ❌ |\n| ✅ Phi-3 Vision | 4.2B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.14219) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FPhi-3-vision-128k-instruct) | [License](https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FPhi-3-mini-4k-instruct\u002Fblob\u002Fmain\u002FLICENSE) | ❌ |\n| ✅ BLIP-2 | 3.6B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2301.12597) | [Hub](https:\u002F\u002Fhuggingface.co\u002FSalesforce\u002Fblip2-opt-2.7b) | MIT | ❌ |\n| ✅ Qwen2-VL | 2B | [Blog](https:\u002F\u002Fqwenlm.github.io\u002Fblog\u002Fqwen2-vl\u002F) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen2-VL-2B-Instruct) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| ✅ Qwen3-VL | 2B\u002F4B\u002F8B | [Blog](https:\u002F\u002Fqwenlm.github.io\u002Fblog\u002Fqwen3-vl\u002F) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3-VL-8B-Instruct) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| ✅ SmolVLM-Instruct | 2B | [Blog](https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fsmolvlm) | [Hub](https:\u002F\u002Fhuggingface.co\u002FHuggingFaceTB\u002FSmolVLM-Instruct) | Apache 2.0 | ✅  |\n\n#### Base Models\n\n| Model | Size | Paper | HF Hub | License | Open |\n|-------|------|-------|---------|----------|------|\n| ✅ SmolLM2 | 135M\u002F360M\u002F1.7B | [Blog](https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fsmollm) | [Hub](https:\u002F\u002Fhuggingface.co\u002FHuggingFaceTB\u002FSmolLM2-135M) | Apache 2.0 | ✅ |\n| ✅ Llama 3.2 | 1B\u002F3B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-3.2-3b) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama3\u002Flicense\u002F) | ❌  |\n| ✅ Llama 3.1 | 8B\u002F70B\u002F405B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-3.1-70b) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama3\u002Flicense\u002F) | ❌  |\n| ✅ GPT-2 | 124M-1.5B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.14165) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fgpt2) | MIT | ✅ |\n| DeepSeek V2 | 7B\u002F13B | [Blog](https:\u002F\u002Fwww.deepseek.com\u002Fblogs\u002Fdeepseek-v2) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fdeepseek-ai\u002Fdeepseek-llm-7b-v2) | [License](https:\u002F\u002Fgithub.com\u002Fdeepseek-ai\u002FDeepSeek-LLM\u002Fblob\u002Fmain\u002FLICENSE-MODEL) | ❌ |\n| Gemma2 | 2B\u002F9B | [Blog](https:\u002F\u002Fai.google.dev\u002Fgemma) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fgoogle\u002Fgemma2-7b) | [License](https:\u002F\u002Fai.google.dev\u002Fgemma\u002Fterms) | ❌ |\n| GPT-J | 6B | [Blog](https:\u002F\u002Fwww.eleuther.ai\u002Fartifacts\u002Fgpt-j) | [Hub](https:\u002F\u002Fhuggingface.co\u002FEleutherAI\u002Fgpt-j-6b) | Apache 2.0 | ✅ |\n| GPT-NeoX | 20B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2204.06745) | [Hub](https:\u002F\u002Fhuggingface.co\u002FEleutherAI\u002Fgpt-neox-20b) | Apache 2.0 | ✅ |\n| Mistral | 7B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.06825) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmistralai\u002FMistral-7B-v0.1) | Apache 2.0 | ❌  |\n| Mixtral | 8x7B\u002F8x22B | [Blog](https:\u002F\u002Fmistral.ai\u002Fnews\u002Fmixtral-of-experts\u002F) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmistralai\u002FMixtral-8x7B-v0.1) | Apache 2.0 | ❌  |\n| MPT | 7B | [Blog](https:\u002F\u002Fwww.mosaicml.com\u002Fblog\u002Fmpt-7b) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmosaicml\u002Fmpt-7b) | Apache 2.0 | ✅ |\n| OLMo | 1B\u002F7B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.00838) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fallenai\u002FOLMo-7B-hf) | Apache 2.0 | ✅ |\n| ✅ Llama 4 Scout | 17B (Activated) 109B (Total) | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-4-Scout-17B-16E) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama4\u002Flicense\u002F) | ❌  |\n\n#### Reasoning Models\n\n| Model | Size | Paper | HF Hub | License | Open |\n|-------|------|-------|---------|----------|------|\n| ✅ gpt-oss | 20B\u002F120B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2508.10925) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fopenai\u002Fgpt-oss-120b) | Apache 2.0 | ❌  |\n| ✅ Qwen3 | 0.6B-32B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2505.09388) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3-32B) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| ✅ Qwen3-Next | 80B-A3B | [Blog](https:\u002F\u002Fqwenlm.github.io\u002Fblog\u002Fqwen3\u002F) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3-Next-80B-A3B) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| Qwen QwQ | 32B | [Blog](https:\u002F\u002Fqwenlm.github.io\u002Fblog\u002Fqwq-32b-preview\u002F) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwQ-32B-Preview) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌ |\n\n#### Code Models\n\n| Model | Size | Paper | HF Hub | License | Open |\n|-------|------|-------|---------|----------|------|\n| ✅ Qwen2.5 Coder | 0.5B-32B | [Blog](https:\u002F\u002Fqwenlm.github.io\u002Fblog\u002Fqwen2.5\u002F) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen2.5-Coder-32B-Instruct) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| DeepSeek Coder | 1.3B-33B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.02954) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fdeepseek-ai\u002Fdeepseek-coder-7b-instruct) | [License](https:\u002F\u002Fgithub.com\u002Fdeepseek-ai\u002FDeepSeek-LLM\u002Fblob\u002Fmain\u002FLICENSE-MODEL) | ❌  |\n| StarCoder 2 | 3B\u002F7B\u002F15B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.19173) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fbigcode\u002Fstarcoder2-15b) | [License](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fbigcode\u002Fbigcode-model-license-agreement) | ✅ |\n\n#### Math Models\n\n| Model | Size | Paper | HF Hub | License | Open |\n|-------|------|-------|---------|----------|------|\n| DeepSeek Math | 7B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.02954) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fdeepseek-ai\u002Fdeepseek-math-7b-instruct) | [License](https:\u002F\u002Fgithub.com\u002Fdeepseek-ai\u002FDeepSeek-LLM\u002Fblob\u002Fmain\u002FLICENSE-MODEL) | ❌  | |\n\n\u003C\u002Fdetails>\n\n## 📖 Documentation\n\nTo learn more about all the platform's capabilities, see the [Oumi documentation](https:\u002F\u002Foumi.ai\u002Fdocs).\n\n## 🤝 Join the Community\n\nOumi is a community-first effort. Whether you are a developer, a researcher, or a non-technical user, all contributions are very welcome!\n\n- To contribute to the `oumi` repository, please check the [`CONTRIBUTING.md`](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002FCONTRIBUTING.md) for guidance on how to contribute to send your first Pull Request.\n- Make sure to join our [Discord community](https:\u002F\u002Fdiscord.gg\u002Foumi) to get help, share your experiences, and contribute to the project!\n- If you are interested in joining one of the community's open-science efforts, check out our [open collaboration](https:\u002F\u002Foumi.ai\u002Fcommunity) page.\n\n## 🙏 Acknowledgements\n\nOumi makes use of [several libraries](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fabout\u002Facknowledgements.html) and tools from the open-source community. We would like to acknowledge and deeply thank the contributors of these projects! ✨ 🌟 💫\n\n## 📝 Citation\n\nIf you find Oumi useful in your research, please consider citing it:\n\n```bibtex\n@software{oumi2025,\n  author = {Oumi Community},\n  title = {Oumi: an Open, End-to-end Platform for Building Large Foundation Models},\n  month = {January},\n  year = {2025},\n  url = {https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi}\n}\n```\n\n## 📜 License\n\nThis project is licensed under the Apache License 2.0. See the [LICENSE](LICENSE) file for details.\n\n[^1]: Open models are defined as models with fully open weights, training code, and data, and a permissive license. See [Open Source Definitions](https:\u002F\u002Fopensource.org\u002Fai) for more information.\n","![Oumi Logo](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Foumi-ai_oumi_readme_0d6331fabdc8.png)\n\n[![文档](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FDocumentation-oumi-blue.svg)](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Findex.html)\n[![博客](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBlog-oumi-blue.svg)](https:\u002F\u002Foumi.ai\u002Fblog)\n[![Twitter](https:\u002F\u002Fimg.shields.io\u002Ftwitter\u002Ffollow\u002FOumi_PBC)](https:\u002F\u002Fx.com\u002FOumi_PBC)\n[![Discord](https:\u002F\u002Fimg.shields.io\u002Fdiscord\u002F1286348126797430814?label=Discord)](https:\u002F\u002Fdiscord.gg\u002Foumi)\n[![PyPI版本](https:\u002F\u002Fbadge.fury.io\u002Fpy\u002Foumi.svg)](https:\u002F\u002Fbadge.fury.io\u002Fpy\u002Foumi)\n[![许可证](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-Apache%202.0-blue.svg)](https:\u002F\u002Fopensource.org\u002Flicenses\u002FApache-2.0)\n[![测试](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Factions\u002Fworkflows\u002Fpretest.yaml\u002Fbadge.svg?branch=main)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Factions\u002Fworkflows\u002Fpretest.yaml)\n[![GPU测试](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Factions\u002Fworkflows\u002Fgpu_tests.yaml\u002Fbadge.svg?branch=main)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Factions\u002Fworkflows\u002Fgpu_tests.yaml)\n[![GitHub仓库星标数](https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Foumi-ai\u002Foumi)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fstargazers)\n[![代码风格：black](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fcode%20style-black-000000.svg)](https:\u002F\u002Fgithub.com\u002Fpsf\u002Fblack)\n[![pre-commit](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fpre--commit-enabled-brightgreen?logo=pre-commit)](https:\u002F\u002Fgithub.com\u002Fpre-commit\u002Fpre-commit)\n[![关于我们](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAbout-oumi-blue.svg)](https:\u002F\u002Foumi.ai)\n\n### 您构建最先进基础模型所需的一切，端到端完成\n\n\u003Cp align=\"center\">\n  \u003Ca href=\"https:\u002F\u002Ftrendshift.io\u002Frepositories\u002F12865\">\n    \u003Cimg alt=\"GitHub趋势\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Foumi-ai_oumi_readme_4a68feb902da.png\" \u002F>\n  \u003C\u002Fa>\n\u003C\u002Fp>\n\n## 🔥 最新消息\n\n- [2026年3月] 升级至Transformers v5、TRL v0.30、vLLM v0.19及veRL v0.7兼容\n- [2026年3月] [MCP集成第一阶段](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F2234)：为MCP服务器支持搭建包框架及依赖\n- [2026年3月] 新增`oumi deploy`命令，用于在fireworks.ai和parasail上部署oumi模型专用推理端点\n- [2026年3月] 增加对Qwen3.5模型系列的支持\n- [2026年3月] 推理引擎获得多项改进：list_models API、错误报告优化等\n- [2026年2月] [Oumi平台与Lambda联合使用，微调并部署4B模型进行用户意图分类预览](https:\u002F\u002Fyoutu.be\u002F0XpfYRpd_FA)\n- [2026年2月] [Lambda与Oumi合作，实现端到端自定义模型开发](https:\u002F\u002Fblog.oumi.ai\u002Fp\u002Flambda-and-oumi-partner-for-end-to)\n- [2025年12月] [Oumi v0.6.0发布](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Freleases\u002Ftag\u002Fv0.6.0)，支持Python 3.13、新增`oumi analyze` CLI命令、TRL 0.26+支持等\n- [2025年12月] [WeMakeDevs AI Agents Assemble黑客马拉松：Oumi关于文本转SQL微调的网络研讨会](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=6wPikqRZ7bQ&t=3203s)\n- [2025年12月] [Oumi与WeMakeDevs共同赞助AI Agents Assemble黑客马拉松，收到超过2000个参赛项目](https:\u002F\u002Fwww.wemakedevs.org\u002Fhackathons\u002Fassemblehack25)\n- [2025年11月] [Oumi v0.5.0发布](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Freleases\u002Ftag\u002Fv0.5)，具备高级数据合成、超参数调优自动化、OpenEnv支持等功能\n- [2025年11月] [使用OpenEnv进行RLVF微调示例笔记本](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi%20-%20OpenEnv%20GRPO%20with%20trl.ipynb)，这是一款由Meta PyTorch团队开发的开源库，用于创建、部署和分发智能体强化学习环境\n- [2025年10月] [Oumi v0.4.1](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Freleases\u002Ftag\u002Fv0.4.1)和[v0.4.2](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Freleases\u002Ftag\u002Fv0.4.2)发布，支持Qwen3-VL和Transformers v4.56、数据合成文档与示例，以及大量错误修复\n\n\u003Cdetails>\n\u003Csummary>更早的更新\u003C\u002Fsummary>\n\n- [2025年9月] [Oumi v0.4.0发布](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Freleases\u002Ftag\u002Fv0.4.0)，支持DeepSpeed、Hugging Face Hub缓存管理工具、KTO\u002FVision DPO训练器\n- [2025年8月] 提供OpenAI的`gpt-oss-20b`和`gpt-oss-120b`的训练与推理支持：[配方在此](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Frecipes\u002Fgpt_oss)\n- [2025年8月] 8月14日网络研讨会——[OpenAI的gpt-oss：去伪存真](https:\u002F\u002Fyoutu.be\u002Fg1PkAV7fXn0)。\n- [2025年8月] [Oumi v0.3.0发布](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Freleases\u002Ftag\u002Fv0.3.0)，包含模型量化（AWQ）、改进的LLM-as-a-Judge API以及自适应推理功能\n- [2025年7月] [Qwen3 235B](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F235b_a22b_together_infer.yaml)的配方\n- [2025年7月] 7月24日网络研讨会：“使用Oumi + Lambda训练最先进的代理LLM”（[观看视频](https:\u002F\u002Fyoutu.be\u002Ff3SU_heBP54)）\n- [2025年6月] [Oumi v0.2.0发布](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Freleases\u002Ftag\u002Fv0.2.0)，支持GRPO微调、新增大量模型支持等\n- [2025年6月] 宣布在NeurIPS2025举办[视觉语言模型数据整理竞赛（DCVLR）](https:\u002F\u002Foumi.ai\u002Fblog\u002Fposts\u002Fannouncing-dcvlr)\n- [2025年6月] 新发布的[Falcon-H1](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Frecipes\u002Ffalcon_h1)和[Falcon-E](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Frecipes\u002Ffalcon_e)模型的训练、推理与评估配方\n- [2025年5月] 对[InternVL3 1B](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Frecipes\u002Fvision\u002Finternvl3)的支持及配方\n- [2025年4月] 新增对Llama 4系列模型的训练与推理支持：Scout（激活17B，总规模109B）和Maverick（激活17B，总规模400B）变体，涵盖完整微调、LoRA及QLoRA配置\n- [2025年4月] [Qwen3模型家族](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Frecipes\u002Fqwen3)的配方\n- [2025年4月] 推出HallOumi：一款最先进的事实核查模型[(技术概述)](https:\u002F\u002Foumi.ai\u002Fblog\u002Fposts\u002Fintroducing-halloumi)\n- [2025年4月] Oumi现支持两款新的视觉语言模型：[Phi4](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Frecipes\u002Fvision\u002Fphi4)和[Qwen 2.5](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b)\n\n\u003C\u002Fdetails>\n\n## 🔎 关于\n\nOumi 是一个完全开源的平台，旨在简化基础模型的整个生命周期——从数据准备、训练，到评估和部署。无论您是在笔记本电脑上进行开发、在集群上启动大规模实验，还是将模型部署到生产环境中，Oumi 都能为您提供所需的工具和工作流。\n\n借助 Oumi，您可以：\n\n- 🚀 使用最先进的技术（SFT、LoRA、QLoRA、GRPO 等）训练和微调参数量从 1000 万到 4050 亿的模型\n- 🤖 同时处理文本模型和多模态模型（Llama、DeepSeek、Qwen、Phi 等）\n- 🔄 利用 LLM 评判员合成并筛选训练数据\n- ⚡️ 使用流行的推理引擎（vLLM、SGLang）高效部署模型\n- 📊 在标准基准上全面评估模型\n- 🌎 可在任何地方运行——从笔记本电脑到集群再到云端（AWS、Azure、GCP、Lambda 等）\n- 🔌 可与开源模型和商业 API（OpenAI、Anthropic、Vertex AI、Together、Parasail 等）集成\n\n所有这些功能都通过一致的 API 实现，具备生产级可靠性，并为您提供研究所需的所有灵活性。\n\n更多信息请访问 [oumi.ai](https:\u002F\u002Foumi.ai\u002Fdocs)，或直接阅读[快速入门指南](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fget_started\u002Fquickstart.html)。\n\n## 🚀 开始使用\n\n| **Notebook** | **在 Colab 中尝试** | **目标** |\n|----------|--------------|-------------|\n| **🎯 入门：概览** | \u003Ca target=\"_blank\" href=\"https:\u002F\u002Fcolab.research.google.com\u002Fgithub\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi - A Tour.ipynb\">\u003Cimg src=\"https:\u002F\u002Fcolab.research.google.com\u002Fassets\u002Fcolab-badge.svg\" alt=\"在 Colab 中打开\"\u002F>\u003C\u002Fa> | 核心功能快速概览：训练、评估、推理和作业管理 |\n| **🔧 模型微调指南** | \u003Ca target=\"_blank\" href=\"https:\u002F\u002Fcolab.research.google.com\u002Fgithub\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi - Finetuning Tutorial.ipynb\">\u003Cimg src=\"https:\u002F\u002Fcolab.research.google.com\u002Fassets\u002Fcolab-badge.svg\" alt=\"在 Colab 中打开\"\u002F>\u003C\u002Fa> | LoRA 微调的端到端指南，包含数据准备、训练和评估 |\n| **📚 模型蒸馏** | \u003Ca target=\"_blank\" href=\"https:\u002F\u002Fcolab.research.google.com\u002Fgithub\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi - Distill a Large Model.ipynb\">\u003Cimg src=\"https:\u002F\u002Fcolab.research.google.com\u002Fassets\u002Fcolab-badge.svg\" alt=\"在 Colab 中打开\"\u002F>\u003C\u002Fa> | 将大型模型蒸馏为更小、更高效的模型的指南 |\n| **📋 模型评估** | \u003Ca target=\"_blank\" href=\"https:\u002F\u002Fcolab.research.google.com\u002Fgithub\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi - Evaluation with Oumi.ipynb\">\u003Cimg src=\"https:\u002F\u002Fcolab.research.google.com\u002Fassets\u002Fcolab-badge.svg\" alt=\"在 Colab 中打开\"\u002F>\u003C\u002Fa> | 使用 Oumi 的评估框架进行全面的模型评估 |\n| **☁️ 远程训练** | \u003Ca target=\"_blank\" href=\"https:\u002F\u002Fcolab.research.google.com\u002Fgithub\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi - Running Jobs Remotely.ipynb\">\u003Cimg src=\"https:\u002F\u002Fcolab.research.google.com\u002Fassets\u002Fcolab-badge.svg\" alt=\"在 Colab 中打开\"\u002F>\u003C\u002Fa> | 在云平台（AWS、Azure、GCP、Lambda 等）上启动并监控训练作业 |\n| **📈 LLM 作为评判员** | \u003Ca target=\"_blank\" href=\"https:\u002F\u002Fcolab.research.google.com\u002Fgithub\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi - Simple Judge.ipynb\">\u003Cimg src=\"https:\u002F\u002Fcolab.research.google.com\u002Fassets\u002Fcolab-badge.svg\" alt=\"在 Colab 中打开\"\u002F>\u003C\u002Fa> | 使用内置评判员过滤和筛选训练数据 |\n\n## 🔧 使用方法\n\n### 安装\n\n选择最适合您的安装方式：\n\n\u003Cdetails open>\n\u003Csummary>\u003Cb>使用 pip（推荐）\u003C\u002Fb>\u003C\u002Fsummary>\n\n```bash\n# 基本安装\nuv pip install oumi\n\n# 带 GPU 支持\nuv pip install 'oumi[gpu]'\n\n# 最新开发版本\nuv pip install git+https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi.git\n```\n\n如果没有 uv？请[安装它](https:\u002F\u002Fdocs.astral.sh\u002Fuv\u002Fgetting-started\u002Finstallation\u002F)，或者改用 `pip`。\n\n\u003C\u002Fdetails>\n\n\u003Cdetails>\n\u003Csummary>\u003Cb>使用 Docker\u003C\u002Fb>\u003C\u002Fsummary>\n\n```bash\n# 拉取最新镜像\ndocker pull ghcr.io\u002Foumi-ai\u002Foumi:latest\n\n# 运行 Oumi 命令\ndocker run --gpus all -it ghcr.io\u002Foumi-ai\u002Foumi:latest oumi --help\n\n# 挂载配置文件进行训练\ndocker run --gpus all -v $(pwd):\u002Fworkspace -it ghcr.io\u002Foumi-ai\u002Foumi:latest \\\n    oumi train --config \u002Fworkspace\u002Fmy_config.yaml\n```\n\n\u003C\u002Fdetails>\n\n\u003Cdetails>\n\u003Csummary>\u003Cb>快速安装脚本（实验性）\u003C\u002Fb>\u003C\u002Fsummary>\n\n无需设置 Python 环境即可试用 Oumi。此脚本会将 Oumi 安装在一个隔离的环境中：\n\n```bash\ncurl -LsSf https:\u002F\u002Foumi.ai\u002Finstall.sh | bash\n```\n\n\u003C\u002Fdetails>\n\n如需更多高级安装选项，请参阅[安装指南](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fget_started\u002Finstallation.html)。\n\n### Oumi CLI\n\n您可以使用 `oumi` 命令快速训练、评估和推理模型，只需选用现有的[配方](\u002Fconfigs\u002Frecipes)之一：\n\n```shell\n# 训练\noumi train -c configs\u002Frecipes\u002Fsmollm\u002Fsft\u002F135m\u002Fquickstart_train.yaml\n\n# 评估\noumi evaluate -c configs\u002Frecipes\u002Fsmollm\u002Fevaluation\u002F135m\u002Fquickstart_eval.yaml\n\n# 推理\noumi infer -c configs\u002Frecipes\u002Fsmollm\u002Finference\u002F135m_infer.yaml --interactive\n```\n\n如需更多高级选项，请参阅[训练](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fuser_guides\u002Ftrain\u002Ftrain.html)、[评估](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fuser_guides\u002Fevaluate\u002Fevaluate.html)、[推理](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fuser_guides\u002Finfer\u002Finfer.html)以及[LLM 作为评判员](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fuser_guides\u002Fjudge\u002Fjudge.html)指南。\n\n### 远程运行作业\n\n您可以通过 `oumi launch` 命令在云平台（AWS、Azure、GCP、Lambda 等）上远程运行作业：\n\n```shell\n# GCP\noumi launch up -c configs\u002Frecipes\u002Fsmollm\u002Fsft\u002F135m\u002Fquickstart_gcp_job.yaml\n\n# AWS\noumi launch up -c configs\u002Frecipes\u002Fsmollm\u002Fsft\u002F135m\u002Fquickstart_gcp_job.yaml --resources.cloud aws\n\n# Azure\noumi launch up -c configs\u002Frecipes\u002Fsmollm\u002Fsft\u002F135m\u002Fquickstart_gcp_job.yaml --resources.cloud azure\n\n# Lambda\noumi launch up -c configs\u002Frecipes\u002Fsmollm\u002Fsft\u002F135m\u002Fquickstart_gcp_job.yaml --resources.cloud lambda\n```\n\n**注意**：Oumi 目前处于\u003Cins>测试版\u003C\u002Fins>阶段，仍在积极开发中。核心功能已稳定，但随着平台不断完善，部分高级功能可能会发生变化。\n\n## 💻 为什么选择 Oumi？\n\n如果您需要一个用于训练、评估或部署模型的综合性平台，Oumi 是一个绝佳的选择。\n\n以下是使 Oumi 脱颖而出的一些关键特性：\n\n- 🔧 **零样板代码**：使用适用于热门模型和工作流的即用型配方，几分钟内即可上手。无需编写训练循环或数据流水线。\n- 🏢 **企业级**：由大规模训练模型的团队构建并验证\n- 🎯 **科研就绪**：非常适合机器学习研究，实验易于复现，且提供灵活的接口以自定义每个组件。\n- 🌐 **广泛的模型支持**：兼容大多数主流模型架构——从小型模型到最大规模的模型，从纯文本模型到多模态模型。\n- 🚀 **SOTA 性能**：原生支持分布式训练技术（FSDP、DeepSpeed、DDP）和优化的推理引擎（vLLM、SGLang）。\n- 🤝 **社区至上**：100% 开源，拥有活跃的社区。无供应商锁定，无附加条件。\n\n## 📚 示例与配方\n\n探索不断增长的、开箱即用的配置集合，适用于最先进的一系列模型和训练工作流：\n\n**注意：** 这些配置并非支持内容的完整列表，而只是帮助您入门的示例。您可以在 oumi 文档中找到更全面的支持 [模型](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fresources\u002Fmodels\u002Fsupported_models.html) 和数据集列表（包括 [监督微调](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fresources\u002Fdatasets\u002Fsft_datasets.html)、[预训练](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fresources\u002Fdatasets\u002Fpretraining_datasets.html)、[偏好优化](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fresources\u002Fdatasets\u002Fpreference_datasets.html) 以及 [视觉-语言微调](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fresources\u002Fdatasets\u002Fvl_sft_datasets.html)）。\n\n### 通义千问家族\n\n| 模型 | 示例配置 |\n|-------|------------------------|\n| Qwen3-Next 80B A3B | [LoRA](\u002Fconfigs\u002Frecipes\u002Fqwen3_next\u002Fsft\u002F80b_a3b_lora\u002Ftrain.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fqwen3_next\u002Finference\u002F80b_a3b_infer.yaml) • [推理（指令版）](\u002Fconfigs\u002Frecipes\u002Fqwen3_next\u002Finference\u002F80b_a3b_instruct_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fqwen3_next\u002Fevaluation\u002F80b_a3b_eval.yaml) |\n| Qwen3 30B A3B | [LoRA](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fsft\u002F30b_a3b_lora\u002Ftrain.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F30b_a3b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fevaluation\u002F30b_a3b_eval.yaml) |\n| Qwen3 32B | [LoRA](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fsft\u002F32b_lora\u002Ftrain.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F32b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fevaluation\u002F32b_eval.yaml) |\n| Qwen3 14B | [LoRA](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fsft\u002F14b_lora\u002Ftrain.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F14b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fevaluation\u002F14b_eval.yaml) |\n| Qwen3 8B | [全量微调](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fsft\u002F8b_full\u002Ftrain.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F8b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fevaluation\u002F8b_eval.yaml) |\n| Qwen3 4B | [全量微调](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fsft\u002F4b_full\u002Ftrain.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F4b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fevaluation\u002F4b_eval.yaml) |\n| Qwen3 1.7B | [全量微调](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fsft\u002F1.7b_full\u002Ftrain.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F1.7b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fevaluation\u002F1.7b_eval.yaml) |\n| Qwen3 0.6B | [全量微调](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fsft\u002F0.6b_full\u002Ftrain.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Finference\u002F0.6b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fqwen3\u002Fevaluation\u002F0.6b_eval.yaml) |\n| QwQ 32B | [全量微调](\u002Fconfigs\u002Frecipes\u002Fqwq\u002Fsft\u002Ffull_train.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fqwq\u002Fsft\u002Flora_train.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fqwq\u002Fsft\u002Fqlora_train.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fqwq\u002Finference\u002Finfer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fqwq\u002Fevaluation\u002Feval.yaml) |\n| Qwen2.5-VL 3B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Fsft\u002Ffull\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Fsft\u002Flora\u002Ftrain.yaml) • [推理（vLLM）](configs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Finference\u002Fvllm_infer.yaml) • [推理](configs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Finference\u002Finfer.yaml) |\n| Qwen2-VL 2B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Fsft\u002Ffull\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Fsft\u002Flora\u002Ftrain.yaml) • [推理（vLLM）](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Finference\u002Fvllm_infer.yaml) • [推理（SGLang）](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Finference\u002Fsglang_infer.yaml) • [推理](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Finference\u002Finfer.yaml) • [评估](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Fevaluation\u002Feval.yaml) |\n\n### 🐋 深势R1家族\n\n| 模型 | 示例配置 |\n|-------|------------------------|\n| DeepSeek R1 671B | [推理（Together AI）](configs\u002Frecipes\u002Fdeepseek_r1\u002Finference\u002F671b_together_infer.yaml) |\n| 精馏Llama 8B | [全量微调](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_llama_8b\u002Ffull_train.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_llama_8b\u002Flora_train.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_llama_8b\u002Fqlora_train.yaml) • [推理](configs\u002Frecipes\u002Fdeepseek_r1\u002Finference\u002Fdistill_llama_8b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fevaluation\u002Fdistill_llama_8b\u002Feval.yaml) |\n| 精馏Llama 70B | [全量微调](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_llama_70b\u002Ffull_train.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_llama_70b\u002Flora_train.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_llama_70b\u002Fqlora_train.yaml) • [推理](configs\u002Frecipes\u002Fdeepseek_r1\u002Finference\u002Fdistill_llama_70b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fevaluation\u002Fdistill_llama_70b\u002Feval.yaml) |\n| 精馏Qwen 1.5B | [全量微调](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_qwen_1_5b\u002Ffull_train.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_qwen_1_5b\u002Flora_train.yaml) • [推理](configs\u002Frecipes\u002Fdeepseek_r1\u002Finference\u002Fdistill_qwen_1_5b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fevaluation\u002Fdistill_qwen_1_5b\u002Feval.yaml) |\n| 精馏Qwen 32B | [LoRA](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fsft\u002Fdistill_qwen_32b\u002Flora_train.yaml) • [推理](configs\u002Frecipes\u002Fdeepseek_r1\u002Finference\u002Fdistill_qwen_32b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fdeepseek_r1\u002Fevaluation\u002Fdistill_qwen_32b\u002Feval.yaml) |\n\n### 🦙 Llama 家族\n\n| 模型 | 示例配置 |\n|-------|------------------------|\n| Llama 4 Scout Instruct 17B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama4\u002Fsft\u002Fscout_instruct_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fllama4\u002Fsft\u002Fscout_instruct_lora\u002Ftrain.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fllama4\u002Fsft\u002Fscout_instruct_qlora\u002Ftrain.yaml) • [推理 (vLLM)](\u002Fconfigs\u002Frecipes\u002Fllama4\u002Finference\u002Fscout_instruct_vllm_infer.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fllama4\u002Finference\u002Fscout_instruct_infer.yaml) • [推理 (Together.ai)](\u002Fconfigs\u002Frecipes\u002Fllama4\u002Finference\u002Fscout_instruct_together_infer.yaml) |\n| Llama 4 Scout 17B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama4\u002Fsft\u002Fscout_base_full\u002Ftrain.yaml)  |\n| Llama 3.1 8B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F8b_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F8b_lora\u002Ftrain.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F8b_qlora\u002Ftrain.yaml) • [预训练](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fpretraining\u002F8b\u002Ftrain.yaml) • [推理 (vLLM)](configs\u002Frecipes\u002Fllama3_1\u002Finference\u002F8b_rvllm_infer.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Finference\u002F8b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fevaluation\u002F8b_eval.yaml) |\n| Llama 3.1 70B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F70b_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F70b_lora\u002Ftrain.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F70b_qlora\u002Ftrain.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Finference\u002F70b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fevaluation\u002F70b_eval.yaml) |\n| Llama 3.1 405B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F405b_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F405b_lora\u002Ftrain.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F405b_qlora\u002Ftrain.yaml) |\n| Llama 3.2 1B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fsft\u002F1b_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fsft\u002F1b_lora\u002Ftrain.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fsft\u002F1b_qlora\u002Ftrain.yaml) • [推理 (vLLM)](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Finference\u002F1b_vllm_infer.yaml) • [推理 (SGLang)](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Finference\u002F1b_sglang_infer.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Finference\u002F1b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fevaluation\u002F1b_eval.yaml) |\n| Llama 3.2 3B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fsft\u002F3b_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fsft\u002F3b_lora\u002Ftrain.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fsft\u002F3b_qlora\u002Ftrain.yaml) • [推理 (vLLM)](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Finference\u002F3b_vllm_infer.yaml) • [推理 (SGLang)](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Finference\u002F3b_sglang_infer.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Finference\u002F3b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fllama3_2\u002Fevaluation\u002F3b_eval.yaml) |\n| Llama 3.3 70B | [FFT](\u002Fconfigs\u002Frecipes\u002Fllama3_3\u002Fsft\u002F70b_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_3\u002Fsft\u002F70b_lora\u002Ftrain.yaml) • [QLoRA](\u002Fconfigs\u002Frecipes\u002Fllama3_3\u002Fsft\u002F70b_qlora\u002Ftrain.yaml) • [推理 (vLLM)](\u002Fconfigs\u002Frecipes\u002Fllama3_3\u002Finference\u002F70b_vllm_infer.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fllama3_3\u002Finference\u002F70b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fllama3_3\u002Fevaluation\u002F70b_eval.yaml) |\n| Llama 3.2 Vision 11B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Fsft\u002F11b_full\u002Ftrain.yaml) • [推理 (vLLM)](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Finference\u002F11b_rvllm_infer.yaml) • [推理 (SGLang)](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Finference\u002F11b_sglang_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Fevaluation\u002F11b_eval.yaml) |\n\n### 🦅 Falcon 家族\n\n| 模型 | 示例配置 |\n|-------|------------------------|\n| [Falcon-H1](https:\u002F\u002Fhuggingface.co\u002Fcollections\u002Ftiiuae\u002Ffalcon-h1-6819f2795bc406da60fab8df) | [FFT](\u002Fconfigs\u002Frecipes\u002Ffalcon_h1\u002Fsft\u002F) • [推理](\u002Fconfigs\u002Frecipes\u002Ffalcon_h1\u002Finference\u002F) • [评估](\u002Fconfigs\u002Frecipes\u002Ffalcon_h1\u002Fevaluation\u002F) |\n| [Falcon-E (BitNet)](https:\u002F\u002Fhuggingface.co\u002Fcollections\u002Ftiiuae\u002Ffalcon-edge-series-6804fd13344d6d8a8fa71130) | [FFT](\u002Fconfigs\u002Frecipes\u002Ffalcon_e\u002Fsft\u002F) • [DPO](\u002Fconfigs\u002Frecipes\u002Ffalcon_e\u002Fdpo\u002F) • [评估](\u002Fconfigs\u002Frecipes\u002Ffalcon_e\u002Fevaluation\u002F) |\n\n### 💎 Gemma 3 家族\n\n| 模型 | 示例配置 |\n|-------|------------------------|\n| Gemma 3 4B Instruct | [FFT](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Fsft\u002F4b_full\u002Ftrain.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Finference\u002F4b_instruct_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Fevaluation\u002F4b\u002Feval.yaml) |\n| Gemma 3 12B Instruct | [LoRA](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Fsft\u002F12b_lora\u002Ftrain.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Finference\u002F12b_instruct_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Fevaluation\u002F12b\u002Feval.yaml) |\n| Gemma 3 27B Instruct | [LoRA](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Fsft\u002F27b_lora\u002Ftrain.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Finference\u002F27b_instruct_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Fgemma3\u002Fevaluation\u002F27b\u002Feval.yaml) |\n\n### 🦉 OLMo 3 家族\n\n| 模型 | 示例配置 |\n|-------|------------------------|\n| OLMo 3 7B Instruct | [FFT](\u002Fconfigs\u002Frecipes\u002Folmo3\u002Fsft\u002F7b_full\u002Ftrain.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Folmo3\u002Finference\u002F7b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Folmo3\u002Fevaluation\u002F7b\u002Feval.yaml) |\n| OLMo 3 32B Instruct | [LoRA](\u002Fconfigs\u002Frecipes\u002Folmo3\u002Fsft\u002F32b_lora\u002Ftrain.yaml) • [推理](\u002Fconfigs\u002Frecipes\u002Folmo3\u002Finference\u002F32b_infer.yaml) • [评估](\u002Fconfigs\u002Frecipes\u002Folmo3\u002Fevaluation\u002F32b\u002Feval.yaml) |\n\n### 🎨 Vision Models\n\n| Model | Example Configurations |\n|-------|------------------------|\n| Llama 3.2 Vision 11B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Fsft\u002F11b_full\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Fsft\u002F11b_lora\u002Ftrain.yaml) • [Inference (vLLM)](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Finference\u002F11b_rvllm_infer.yaml) • [Inference (SGLang)](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Finference\u002F11b_sglang_infer.yaml) • [Evaluation](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllama3_2_vision\u002Fevaluation\u002F11b_eval.yaml) |\n| LLaVA 7B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllava_7b\u002Fsft\u002Ftrain.yaml) • [Inference (vLLM)](configs\u002Frecipes\u002Fvision\u002Fllava_7b\u002Finference\u002Fvllm_infer.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fllava_7b\u002Finference\u002Finfer.yaml) |\n| Phi3 Vision 4.2B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fphi3\u002Fsft\u002Ffull\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fphi3\u002Fsft\u002Flora\u002Ftrain.yaml) • [Inference (vLLM)](configs\u002Frecipes\u002Fvision\u002Fphi3\u002Finference\u002Fvllm_infer.yaml) |\n| Phi4 Vision 5.6B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fphi4\u002Fsft\u002Ffull\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fphi4\u002Fsft\u002Flora\u002Ftrain.yaml) • [Inference (vLLM)](configs\u002Frecipes\u002Fvision\u002Fphi4\u002Finference\u002Fvllm_infer.yaml) • [Inference](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fphi4\u002Finference\u002Finfer.yaml) |\n| Qwen2-VL 2B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Fsft\u002Ffull\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Fsft\u002Flora\u002Ftrain.yaml) • [Inference (vLLM)](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Finference\u002Fvllm_infer.yaml) • [Inference (SGLang)](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Finference\u002Fsglang_infer.yaml) • [Inference](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Finference\u002Finfer.yaml) • [Evaluation](configs\u002Frecipes\u002Fvision\u002Fqwen2_vl_2b\u002Fevaluation\u002Feval.yaml) |\n| Qwen3-VL 2B | [Inference](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen3_vl_2b\u002Finference\u002Finfer.yaml) |\n| Qwen3-VL 4B | [Inference](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen3_vl_4b\u002Finference\u002Finfer.yaml) |\n| Qwen3-VL 8B | [Inference](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen3_vl_8b\u002Finference\u002Finfer.yaml) |\n| Qwen2.5-VL 3B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Fsft\u002Ffull\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Fsft\u002Flora\u002Ftrain.yaml)• [Inference (vLLM)](configs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Finference\u002Fvllm_infer.yaml) • [Inference](configs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Finference\u002Finfer.yaml) |\n| SmolVLM-Instruct 2B | [SFT](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fsmolvlm\u002Fsft\u002Ffull\u002Ftrain.yaml) • [LoRA](\u002Fconfigs\u002Frecipes\u002Fvision\u002Fsmolvlm\u002Fsft\u002Flora\u002Ftrain.yaml) |\n\n### 🔍 Even more options\n\nThis section lists all the language models that can be used with Oumi. Thanks to the integration with the [🤗 Transformers](https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Ftransformers) library, you can easily use any of these models for training, evaluation, or inference.\n\nModels prefixed with a checkmark (✅) have been thoroughly tested and validated by the Oumi community, with ready-to-use recipes available in the [configs\u002Frecipes](configs\u002Frecipes) directory.\n\n\u003Cdetails>\n\u003Csummary>📋 Click to see more supported models\u003C\u002Fsummary>\n\n#### Instruct Models\n\n| Model | Size | Paper | HF Hub  | License  | Open [^1] |\n|-------|------|-------|---------|----------|------|\n| ✅ SmolLM-Instruct | 135M\u002F360M\u002F1.7B | [Blog](https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fsmollm) | [Hub](https:\u002F\u002Fhuggingface.co\u002FHuggingFaceTB\u002FSmolLM-135M-Instruct) | Apache 2.0 | ✅ |\n| ✅ DeepSeek R1 Family | 1.5B\u002F8B\u002F32B\u002F70B\u002F671B | [Blog](https:\u002F\u002Fapi-docs.deepseek.com\u002Fnews\u002Fnews250120) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fdeepseek-ai\u002FDeepSeek-R1) | MIT | ❌ |\n| ✅ Llama 3.1 Instruct | 8B\u002F70B\u002F405B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-3.1-70b-instruct) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama3\u002Flicense\u002F) | ❌  |\n| ✅ Llama 3.2 Instruct | 1B\u002F3B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-3.2-3b-instruct) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama3\u002Flicense\u002F) | ❌  |\n| ✅ Llama 3.3 Instruct | 70B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-3.3-70b-instruct) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama3\u002Flicense\u002F) | ❌  |\n| ✅ Phi-3.5-Instruct | 4B\u002F14B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.14219) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FPhi-3.5-mini-instruct) | [License](https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FPhi-3-mini-4k-instruct\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| ✅ Qwen3 | 0.6B-32B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2505.09388) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3-32B) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| Qwen2.5-Instruct | 0.5B-70B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.16609) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen2.5-7B-Instruct) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| OLMo 2 Instruct | 7B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.00838) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fallenai\u002FOLMo-2-1124-7B) | Apache 2.0 | ✅ |\n| ✅ OLMo 3 Instruct | 7B\u002F32B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.00838) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fallenai\u002FOLMo-3-7B-Instruct) | Apache 2.0 | ✅ |\n| MPT-Instruct | 7B | [Blog](https:\u002F\u002Fwww.mosaicml.com\u002Fblog\u002Fmpt-7b) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmosaicml\u002Fmpt-7b-instruct) | Apache 2.0 | ✅ |\n| Command R | 35B\u002F104B | [Blog](https:\u002F\u002Fcohere.com\u002Fblog\u002Fcommand-r7b) | [Hub](https:\u002F\u002Fhuggingface.co\u002FCohereForAI\u002Fc4ai-command-r-plus) | [License](https:\u002F\u002Fcohere.com\u002Fc4ai-cc-by-nc-license) | ❌ |\n| Granite-3.1-Instruct | 2B\u002F8B | [Paper](https:\u002F\u002Fgithub.com\u002Fibm-granite\u002Fgranite-3.0-language-models\u002Fblob\u002Fmain\u002Fpaper.pdf) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fibm-granite\u002Fgranite-3.1-8b-instruct) | Apache 2.0 | ❌ |\n| Gemma 2 Instruct | 2B\u002F9B | [Blog](https:\u002F\u002Fai.google.dev\u002Fgemma) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fgoogle\u002Fgemma-2-2b-it) | [License](https:\u002F\u002Fai.google.dev\u002Fgemma\u002Fterms) | ❌ |\n| ✅ Gemma 3 Instruct | 4B\u002F12B\u002F27B | [Blog](https:\u002F\u002Fai.google.dev\u002Fgemma) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fgoogle\u002Fgemma-3-27b-it) | [License](https:\u002F\u002Fai.google.dev\u002Fgemma\u002Fterms) | ❌ |\n| DBRX-Instruct | 130B MoE | [Blog](https:\u002F\u002Fwww.databricks.com\u002Fblog\u002Fintroducing-dbrx-new-state-art-open-llm) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fdatabricks\u002Fdbrx-instruct) | Apache 2.0 | ❌ |\n| Falcon-Instruct | 7B\u002F40B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2306.01116) | [Hub](https:\u002F\u002Fhuggingface.co\u002Ftiiuae\u002Ffalcon-7b-instruct) | Apache 2.0 | ❌  |\n| ✅ Llama 4 Scout Instruct | 17B (Activated) 109B (Total) | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-4-Scout-17B-16E-Instruct) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama4\u002Flicense\u002F) | ❌  |\n| ✅ Llama 4 Maverick Instruct | 17B (Activated) 400B (Total) | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-4-Maverick-17B-128E-Instruct) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama4\u002Flicense\u002F) | ❌  |\n\n#### Vision-Language Models\n\n| Model | Size | Paper | HF Hub | License | Open |\n|-------|------|-------|---------|----------|------|\n| ✅ Llama 3.2 Vision | 11B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-3.2-11b-vision) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama3\u002Flicense\u002F) | ❌  |\n| ✅ LLaVA-1.5 | 7B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.03744) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fllava-hf\u002Fllava-1.5-7b-hf) | [License](https:\u002F\u002Fai.meta.com\u002Fllama\u002Flicense) | ❌ |\n| ✅ Phi-3 Vision | 4.2B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.14219) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FPhi-3-vision-128k-instruct) | [License](https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FPhi-3-mini-4k-instruct\u002Fblob\u002Fmain\u002FLICENSE) | ❌ |\n| ✅ BLIP-2 | 3.6B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2301.12597) | [Hub](https:\u002F\u002Fhuggingface.co\u002FSalesforce\u002Fblip2-opt-2.7b) | MIT | ❌ |\n| ✅ Qwen2-VL | 2B | [Blog](https:\u002F\u002Fqwenlm.github.io\u002Fblog\u002Fqwen2-vl\u002F) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen2-VL-2B-Instruct) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| ✅ Qwen3-VL | 2B\u002F4B\u002F8B | [Blog](https:\u002F\u002Fqwenlm.github.io\u002Fblog\u002Fqwen3-vl\u002F) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3-VL-8B-Instruct) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| ✅ SmolVLM-Instruct | 2B | [Blog](https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fsmolvlm) | [Hub](https:\u002F\u002Fhuggingface.co\u002FHuggingFaceTB\u002FSmolVLM-Instruct) | Apache 2.0 | ✅  |\n\n#### Base Models\n\n| Model | Size | Paper | HF Hub | License | Open |\n|-------|------|-------|---------|----------|------|\n| ✅ SmolLM2 | 135M\u002F360M\u002F1.7B | [Blog](https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fsmollm) | [Hub](https:\u002F\u002Fhuggingface.co\u002FHuggingFaceTB\u002FSmolLM2-135M) | Apache 2.0 | ✅ |\n| ✅ Llama 3.2 | 1B\u002F3B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-3.2-3b) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama3\u002Flicense\u002F) | ❌  |\n| ✅ Llama 3.1 | 8B\u002F70B\u002F405B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-3.1-70b) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama3\u002Flicense\u002F) | ❌  |\n| ✅ GPT-2 | 124M-1.5B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.14165) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fgpt2) | MIT | ✅ |\n| DeepSeek V2 | 7B\u002F13B | [Blog](https:\u002F\u002Fwww.deepseek.com\u002Fblogs\u002Fdeepseek-v2) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fdeepseek-ai\u002Fdeepseek-llm-7b-v2) | [License](https:\u002F\u002Fgithub.com\u002Fdeepseek-ai\u002FDeepSeek-LLM\u002Fblob\u002Fmain\u002FLICENSE-MODEL) | ❌ |\n| Gemma2 | 2B\u002F9B | [Blog](https:\u002F\u002Fai.google.dev\u002Fgemma) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fgoogle\u002Fgemma2-7b) | [License](https:\u002F\u002Fai.google.dev\u002Fgemma\u002Fterms) | ❌ |\n| GPT-J | 6B | [Blog](https:\u002F\u002Fwww.eleuther.ai\u002Fartifacts\u002Fgpt-j) | [Hub](https:\u002F\u002Fhuggingface.co\u002FEleutherAI\u002Fgpt-j-6b) | Apache 2.0 | ✅ |\n| GPT-NeoX | 20B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2204.06745) | [Hub](https:\u002F\u002Fhuggingface.co\u002FEleutherAI\u002Fgpt-neox-20b) | Apache 2.0 | ✅ |\n| Mistral | 7B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.06825) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmistralai\u002FMistral-7B-v0.1) | Apache 2.0 | ❌  |\n| Mixtral | 8x7B\u002F8x22B | [Blog](https:\u002F\u002Fmistral.ai\u002Fnews\u002Fmixtral-of-experts\u002F) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmistralai\u002FMixtral-8x7B-v0.1) | Apache 2.0 | ❌  |\n| MPT | 7B | [Blog](https:\u002F\u002Fwww.mosaicml.com\u002Fblog\u002Fmpt-7b) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmosaicml\u002Fmpt-7b) | Apache 2.0 | ✅ |\n| OLMo | 1B\u002F7B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.00838) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fallenai\u002FOLMo-7B-hf) | Apache 2.0 | ✅ |\n| ✅ Llama 4 Scout | 17B (Activated) 109B (Total) | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.21783) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fmeta-llama\u002FLlama-4-Scout-17B-16E) | [License](https:\u002F\u002Fllama.meta.com\u002Fllama4\u002Flicense\u002F) | ❌  |\n\n#### Reasoning Models\n\n| Model | Size | Paper | HF Hub | License | Open |\n|-------|------|-------|---------|----------|------|\n| ✅ gpt-oss | 20B\u002F120B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2508.10925) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fopenai\u002Fgpt-oss-120b) | Apache 2.0 | ❌  |\n| ✅ Qwen3 | 0.6B-32B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2505.09388) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3-32B) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| ✅ Qwen3-Next | 80B-A3B | [Blog](https:\u002F\u002Fqwenlm.github.io\u002Fblog\u002Fqwen3\u002F) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3-Next-80B-A3B) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| Qwen QwQ | 32B | [Blog](https:\u002F\u002Fqwenlm.github.io\u002Fblog\u002Fqwq-32b-preview\u002F) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwQ-32B-Preview) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌ |\n\n#### Code Models\n\n| Model | Size | Paper | HF Hub | License | Open |\n|-------|------|-------|---------|----------|------|\n| ✅ Qwen2.5 Coder | 0.5B-32B | [Blog](https:\u002F\u002Fqwenlm.github.io\u002Fblog\u002Fqwen2.5\u002F) | [Hub](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen2.5-Coder-32B-Instruct) | [License](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen\u002Fblob\u002Fmain\u002FLICENSE) | ❌  |\n| DeepSeek Coder | 1.3B-33B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.02954) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fdeepseek-ai\u002Fdeepseek-coder-7b-instruct) | [License](https:\u002F\u002Fgithub.com\u002Fdeepseek-ai\u002FDeepSeek-LLM\u002Fblob\u002Fmain\u002FLICENSE-MODEL) | ❌  |\n| StarCoder 2 | 3B\u002F7B\u002F15B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.19173) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fbigcode\u002Fstarcoder2-15b) | [License](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fbigcode\u002Fbigcode-model-license-agreement) | ✅ |\n\n#### Math Models\n\n| Model | Size | Paper | HF Hub | License | Open |\n|-------|------|-------|---------|----------|------|\n| DeepSeek Math | 7B | [Paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.02954) | [Hub](https:\u002F\u002Fhuggingface.co\u002Fdeepseek-ai\u002Fdeepseek-math-7b-instruct) | [License](https:\u002F\u002Fgithub.com\u002Fdeepseek-ai\u002FDeepSeek-LLM\u002Fblob\u002Fmain\u002FLICENSE-MODEL) | ❌  | |\n\n\u003C\u002Fdetails>\n\n\n\n## 📖 Documentation\n\nTo learn more about all the platform's capabilities, see the [Oumi documentation](https:\u002F\u002Foumi.ai\u002Fdocs).\n\n## 🤝 Join the Community\n\nOumi is a community-first effort. Whether you are a developer, a researcher, or a non-technical user, all contributions are very welcome!\n\n- To contribute to the `oumi` repository, please check the [`CONTRIBUTING.md`](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002FCONTRIBUTING.md) for guidance on how to contribute to send your first Pull Request.\n- Make sure to join our [Discord community](https:\u002F\u002Fdiscord.gg\u002Foumi) to get help, share your experiences, and contribute to the project!\n- If you are interested in joining one of the community's open-science efforts, check out our [open collaboration](https:\u002F\u002Foumi.ai\u002Fcommunity) page.\n\n## 🙏 Acknowledgements\n\nOumi makes use of [several libraries](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fabout\u002Facknowledgements.html) and tools from the open-source community. We would like to acknowledge and deeply thank the contributors of these projects! ✨ 🌟 💫\n\n## 📝 Citation\n\nIf you find Oumi useful in your research, please consider citing it:\n\n```bibtex\n@software{oumi2025,\n  author = {Oumi Community},\n  title = {Oumi: an Open, End-to-end Platform for Building Large Foundation Models},\n  month = {January},\n  year = {2025},\n  url = {https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi}\n}\n```\n\n## 📜 License\n\nThis project is licensed under the Apache License 2.0. See the [LICENSE](LICENSE) file for details.\n\n[^1]: Open models are defined as models with fully open weights, training code, and data, and a permissive license. See [Open Source Definitions](https:\u002F\u002Fopensource.org\u002Fai) for more information.","# Oumi 快速上手指南\n\nOumi 是一个完全开源的平台，旨在简化基础模型的全生命周期管理——从数据准备、训练、评估到部署。它支持从笔记本电脑到大规模集群的多种环境，并提供统一的 API 来训练和微调从 10M 到 405B 参数的模型。\n\n## 环境准备\n\n在开始之前，请确保您的开发环境满足以下要求：\n\n*   **操作系统**: Linux 或 macOS (Windows 用户建议使用 WSL2 或 Docker)。\n*   **Python 版本**: 推荐 Python 3.10 - 3.13。\n*   **硬件要求**:\n    *   **基础使用**: 任意现代 CPU 即可运行推理和小型实验。\n    *   **GPU 训练\u002F推理**: 需要 NVIDIA GPU 并安装对应的 CUDA 驱动。\n*   **前置依赖**:\n    *   推荐使用 [`uv`](https:\u002F\u002Fdocs.astral.sh\u002Fuv\u002F) 进行包管理（速度更快），若未安装也可使用标准的 `pip`。\n    *   若需远程启动云任务，需配置对应云平台（AWS, GCP, Azure, Lambda 等）的凭证。\n\n> **国内开发者提示**：由于 PyPI 源在国外可能访问较慢，建议在使用 `pip` 或 `uv` 时配置国内镜像源（如清华源、阿里源）。\n> *   `uv` 配置示例：`export UV_INDEX_URL=https:\u002F\u002Fpypi.tuna.tsinghua.edu.cn\u002Fsimple`\n> *   `pip` 配置示例：`pip install ... -i https:\u002F\u002Fpypi.tuna.tsinghua.edu.cn\u002Fsimple`\n\n## 安装步骤\n\n您可以选择以下任一方式进行安装：\n\n### 方式一：使用 uv 安装（推荐）\n\n这是最快且最可靠的安装方式。\n\n```bash\n# 基础安装\nuv pip install oumi\n\n# 如果需要 GPU 支持（包含 torch, vLLM 等深度学习依赖）\nuv pip install 'oumi[gpu]'\n\n# 或者安装最新开发版本\nuv pip install git+https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi.git\n```\n\n*如果没有安装 `uv`，可以先安装它，或者将上述命令中的 `uv pip` 替换为 `pip`。*\n\n### 方式二：使用 Docker\n\n如果您希望避免配置本地 Python 环境，可以直接使用官方 Docker 镜像：\n\n```bash\n# 拉取最新镜像\ndocker pull ghcr.io\u002Foumi-ai\u002Foumi:latest\n\n# 运行测试命令\ndocker run --gpus all -it ghcr.io\u002Foumi-ai\u002Foumi:latest oumi --help\n\n# 挂载本地配置文件并运行训练\ndocker run --gpus all -v $(pwd):\u002Fworkspace -it ghcr.io\u002Foumi-ai\u002Foumi:latest \\\n    oumi train --config \u002Fworkspace\u002Fmy_config.yaml\n```\n\n### 方式三：一键安装脚本（实验性）\n\n在不设置复杂 Python 环境的情况下快速尝试：\n\n```bash\ncurl -LsSf https:\u002F\u002Foumi.ai\u002Finstall.sh | bash\n```\n\n## 基本使用\n\n安装完成后，您可以直接使用 `oumi` 命令行工具调用预置的配方（Recipes）进行训练、评估和推理。以下以 `smollm` 模型为例：\n\n### 1. 模型训练 (Training)\n\n使用预置配置文件启动监督微调 (SFT)：\n\n```shell\noumi train -c configs\u002Frecipes\u002Fsmollm\u002Fsft\u002F135m\u002Fquickstart_train.yaml\n```\n\n### 2. 模型评估 (Evaluation)\n\n在标准基准测试上评估模型性能：\n\n```shell\noumi evaluate -c configs\u002Frecipes\u002Fsmollm\u002Fevaluation\u002F135m\u002Fquickstart_eval.yaml\n```\n\n### 3. 模型推理 (Inference)\n\n启动交互式对话界面进行测试：\n\n```shell\noumi infer -c configs\u002Frecipes\u002Fsmollm\u002Finference\u002F135m_infer.yaml --interactive\n```\n\n### 4. 远程云端任务 (可选)\n\n如果您需要在云端集群（如 AWS, GCP, Lambda）上运行任务，可以使用 `oumi launch` 命令：\n\n```shell\n# 在 GCP 上启动任务\noumi launch up -c configs\u002Frecipes\u002Fsmollm\u002Fsft\u002F135m\u002Fquickstart_gcp_job.yaml\n\n# 指定在 AWS 上启动\noumi launch up -c configs\u002Frecipes\u002Fsmollm\u002Fsft\u002F135m\u002Fquickstart_gcp_job.yaml --resources.cloud aws\n```\n\n---\n**下一步建议**：\n对于更详细的功能演示（如 LoRA 微调、模型蒸馏、数据合成等），建议访问 Oumi 官方文档或直接在 Google Colab 中运行其提供的 [入门教程 Notebook](https:\u002F\u002Fcolab.research.google.com\u002Fgithub\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi%20-%20A%20Tour.ipynb)。","某电商初创公司的算法团队急需构建一个专属的“智能售后客服模型”，以自动处理海量的用户退货与换货请求，但团队仅有两名工程师且算力资源有限。\n\n### 没有 oumi 时\n- **环境配置繁琐**：工程师需手动拼接 Hugging Face Transformers、TRL 和 vLLM 等库的特定版本，常因依赖冲突导致数天的环境调试时间浪费。\n- **微调流程割裂**：数据预处理、模型训练（如 Qwen3 或 DeepSeek-R1）与后续评估分散在不同脚本中，缺乏统一标准，难以复现实验结果。\n- **部署门槛极高**：将训练好的模型转化为高并发推理服务需要编写复杂的后端代码和 Docker 配置，小团队无力承担运维成本。\n- **模型选型困难**：面对众多开源多模态模型（VLM），缺乏便捷工具快速验证哪个模型最适合处理“用户上传的破损商品图片”场景。\n\n### 使用 oumi 后\n- **一键式全链路管理**：oumi 提供端到端解决方案，统一兼容主流框架，工程师只需一条命令即可完成从环境准备到依赖安装的全过程。\n- **标准化工作流**：通过配置文件即可驱动数据合成、超参数自动调优及模型微调，轻松复现针对售后场景优化的 Qwen3.5 模型效果。\n- **极速云端部署**：利用 `oumi deploy` 命令，直接将微调后的模型发布到 Fireworks.ai 或 Parasail 等专用推理端点，分钟级上线服务。\n- **高效模型评估**：内置评估工具可快速对比不同开源模型在“图文理解”任务上的表现，迅速锁定最适合处理退货图片的模型架构。\n\noumi 让小型团队也能像大厂一样，以极低的工程成本实现从开源模型选择、定制微调到生产级部署的闭环落地。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Foumi-ai_oumi_87ee2eaa.png","oumi-ai","Oumi","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002Foumi-ai_c4d95975.png","Making frontier AI truly open for all",null,"contact@oumi.ai","Oumi_PBC","https:\u002F\u002Foumi.ai\u002F","https:\u002F\u002Fgithub.com\u002Foumi-ai",[83,87,91,95,99,103],{"name":84,"color":85,"percentage":86},"Python","#3572A5",89.1,{"name":88,"color":89,"percentage":90},"Jupyter Notebook","#DA5B0B",9.4,{"name":92,"color":93,"percentage":94},"Shell","#89e051",1.2,{"name":96,"color":97,"percentage":98},"Jinja","#a52a22",0.2,{"name":100,"color":101,"percentage":102},"Makefile","#427819",0.1,{"name":104,"color":105,"percentage":106},"Dockerfile","#384d54",0,9161,742,"2026-04-07T00:11:32","Apache-2.0","Linux, macOS","训练和大规模推理需要 NVIDIA GPU（支持 CUDA），具体显存取决于模型大小（从 10M 到 405B 参数）；支持 CPU 运行小型任务或推理。","未说明（取决于模型规模，大模型训练需高内存）",{"notes":115,"python":116,"dependencies":117},"推荐使用 'uv' 或 'pip' 安装，可通过 'oumi[gpu]' 获取 GPU 支持版本。支持多种云厂商（AWS, Azure, GCP, Lambda）远程任务调度。支持多种推理后端（vLLM, SGLang）。提供 Docker 镜像以便快速部署。支持从笔记本到集群的多种运行环境。","3.9+ (v0.6.0 起支持 Python 3.13)",[118,119,120,121,122,123,124,125,126,127],"torch","transformers>=5.0","trl>=0.30","vllm>=0.19","deepspeed","accelerate","peft","datasets","omegaconf","rich",[35,14,129],"其他",[131,132,133,134,135,136,137,138,139,140,141,142],"dpo","evaluation","fine-tuning","inference","llama","llms","sft","vlms","gpt-oss","gpt-oss-120b","gpt-oss-20b","slms","2026-03-27T02:49:30.150509","2026-04-07T13:29:03.518505",[],[147,152,157,162,167,172,177,182,187,192,197,202,207,212,217,222,227,232,237,242],{"id":148,"version":149,"summary_zh":150,"released_at":151},135928,"v0.7","# Oumi 0.7 版本发布\n\n## ✨ 亮点\n\n本次发布带来了重大的平台升级（Python 3.14、PyTorch 2.9）、新的推理引擎、基于规则的评估裁判，以及 CLI 和文档的重大改进。\n\n---\n\n\n\n## 🚀 新特性\n\n### 推理\n\n- **Fireworks 推理引擎** - Fireworks AI 的新后端 (#2158)\n\n\n```yaml\n# Fireworks 示例（需设置 FIREWORKS_API_KEY 环境变量）\nmodel:\n  model_name: \"accounts\u002Ffireworks\u002Fmodels\u002Fllama4-maverick-instruct-basic\"\nengine: FIREWORKS\n```\n\n\n```bash\noumi infer -i -c configs\u002Frecipes\u002Fllama4\u002Finference\u002Fmaverick_instruct_fireworks_infer.yaml\n```  \n- **OpenRouter 推理引擎** - OpenRouter 的新后端 (#2168)\n\n```yaml\n# OpenRouter 示例（需设置 OPENROUTER_API_KEY 环境变量）\nmodel:\n  model_name: \"anthropic\u002Fclaude-sonnet-4.5\"\nengine: OPENROUTER\n```\n\n\n```bash\n# 通过 CLI 使用\noumi infer -i -c configs\u002Fapis\u002Fopenrouter\u002Finfer_claude_4_5_sonnet.yaml\n```  \n\n- **加载进度条** - 推理操作期间的可视化反馈 (#2085)\n- **预训练自定义模型支持** - 加载您自己的预训练模型 (#2044)\n\n### 评估\n\n- **基于规则的裁判** - 具有 CLI 集成和示例的确定性评估裁判 (#2119, #2171)\n\n```yaml\n# configs\u002Fprojects\u002Fjudges\u002Frule_based\u002Fregex_match_phone.yaml\njudge_params:\n  prompt_template: \"{response}\"\n\nrule_judge_params:\n  rule_type: \"正则表达式\"\n  input_fields:\n    - \"response\"\n  rule_config:\n    pattern: \"\\\\d{3}-\\\\d{4}\"\n    input_field: \"response\"\n    match_mode: \"搜索\"\n    inverse: false\n```\n\n\n```bash\noumi judge dataset -c regex-match-phone --input data\u002Fjudge_input.jsonl\n```  \n\n### 训练\n\n- **指标日志回调** - 将训练指标记录到磁盘 (#2140)\n- **每项奖励函数配置** - 新增 `reward_function_kwargs` 支持 (#2143)\n\n```yaml\ntrainer_type: TRL_GRPO\n\nreward_functions:\n  - rubric_reward\n  - gsm8k\n\nreward_function_kwargs:\n  rubric_reward:\n    judge_panel_path: \"configs\u002Fprojects\u002Fjudges\u002Frubric_panel.yaml\"\n  gsm8k:\n    strict: true\n```  \n\n### 数据与合成\n\n- **XLSX 和 DOCX 支持** - 合成和数据集的新格式 (#2148)\n- **少量样本采样** - 在合成过程中从来源中采样少量示例 (#2151)\n- **批量 AttributeSynthesizer** - 批量处理支持 (#2181)\n- **RaR 数据集** - 新的数据集和基础评分数据集类 (#2144)\n\n### 基础设施\n\n- **Nebius 云提供商** - 新的云选项 (#2179)\n- **Kubernetes Skypilot 支持** - 添加了 k8s 依赖 (#2124)\n- **ARM Docker 支持** - 启用了 ARM 构建并提供了实用工具 (#2141)\n- **一键安装程序** - 新的 `install.sh` 脚本 (#2155)\n\n```bash\n# 基本安装\ncurl -LsSf https:\u002F\u002Foumi.ai\u002Finstall.sh | bash\n\n# 带 GPU 支持\ncurl -LsSf https:\u002F\u002Foumi.ai\u002Finstall.sh | bash -s -- --gpu\n\n# 指定 Python 版本\ncurl -LsSf https:\u002F\u002Foumi.ai\u002Finstall.sh | bash -s -- --python 3.12\n```  \n\n- **遥测** - 通过 PostHog 提供可选的使用情况分析 (#2145)\n\n---","2026-01-29T19:59:20",{"id":153,"version":154,"summary_zh":155,"released_at":156},135929,"v0.6.0","# Oumi v0.6.0 更改日志\n\n我们很高兴地宣布 **Oumi v0.6.0**！此版本带来了 **Python 3.13 支持**、功能强大的全新 **数据集分析 CLI 工具**、用于偏好学习的 **TRL GOLD 训练器**，以及 **一流的 Kubernetes 部署支持**。\n\n---\n\n## 亮点\n\n### Python 3.13 支持\nOumi 现已正式支持 **Python 3.13**，让您能够充分利用最新的 Python 性能改进和新特性。  \n*(#2092)*\n\n---\n\n### 新的 `oumi analyze` CLI 命令\n理解您的训练数据现在变得更加简单。全新的 `oumi analyze` 命令允许您直接从命令行检查和分析数据集——无需编写任何代码。\n\n```bash\n# 分析本地数据集\noumi analyze -c configs\u002Fexamples\u002Fanalyze\u002Fanalyze.yaml\n```\n\n```bash\n# 将结果导出为不同格式\noumi analyze -c configs\u002Fexamples\u002Fanalyze\u002Fanalyze.yaml --format parquet --output .\u002Fmy_results\n```\n\n只需创建一个简单的配置文件，即可分析任何 HuggingFace 数据集：\n\n```yaml\n# hf_analyze.yaml\ndataset_name: argilla\u002Fdatabricks-dolly-15k-curated-en\nsplit: train\nsample_count: 1000\nanalyzers:\n  - id: length\n```\n\n更多详细信息请参阅分析文档。  \n*(#2069, #2071)*\n\n---\n\n### TRL GOLD 训练器\n我们新增了对 TRL 提供的 **[GOLD（基于示范的广义在线学习](https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.02743)** 训练器的支持。GOLD 是一种在线偏好学习算法，它在 DPO 的基础上进行了改进，在训练过程中实时生成响应，从而以更小的分布偏移实现更好的对齐效果。\n\n```bash\n# 使用示例配置运行 GOLD 训练\noumi train -c configs\u002Fexamples\u002Fgold\u002Ftrain.yaml\n```\n\n或者在您自己的训练配置中进行设置：\n\n```yaml\ntraining:\n  trainer_type: \"TRL_GOLD\"\n  gold:\n    teacher_model_name_or_path: \"HuggingFaceTB\u002FSmolLM2-360M-Instruct\"\n    temperature: 0.9\n    max_completion_length: 512\n    lmbda: 0.5  # 50% 在策略，50% 离策略\n```\n\n这需要 **TRL 0.26+**，该版本现已成为默认版本。  \n*(#2095, #2097)*\n\n---\n\n### 代码评估裁判\n新增了专门用于评估代码质量的 **LLM 作为裁判** 评估器。这些裁判可以评估生成代码的正确性、风格、安全性以及其他软件工程最佳实践——非常适合用于评估编码助手和代码生成模型。\n\n感谢 **@N-45div** 的贡献！  \n*(#2087)*\n\n---\n\n### Kubernetes 部署\n现在您可以在 Kubernetes 集群上部署 Oumi 训练任务。\n\n#### 选项 1：使用 SkyPilot（本版本新增）\n```yaml\n# k8s_job.yaml\nname: my-training-job\nresources:\n  cloud: k8s\n  accelerators: \"A100:1\"\nrun: |\n  oumi train -c configs\u002Frecipes\u002Fllama3_1\u002Fsft\u002F8b_lora\u002Ftrain.yaml\n```\n\n```bash\noumi launch up -c k8s_job.yaml --cluster my-k8s-cluster\n```\n\n#### 选项 2：直接使用 kubectl 部署\n对于现有的 Kubernetes 集群，您可以直接使用 kubectl 部署 Oumi。请参阅 [Kubernetes 部署指南](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fu","2025-12-17T21:26:30",{"id":158,"version":159,"summary_zh":160,"released_at":161},135930,"v0.5.0","# **Oumi v0.5.0 发行说明**\n\n我们很高兴地宣布 Oumi v0.5.0 正式发布，此次版本新增了超参数调优功能、扩展了推理选项，并增强了启动器的功能。\n\n\n## **🚀 主要特性**\n\n\n### **数据合成模块**\n\n\n\n* 推出 `oumi synth` —— 一个强大的数据合成模块，可利用大语言模型自动生成高质量的训练数据集（#1965）\n* **基于模板的生成**：通过控制难度、风格和领域等属性，创建多样化的数据集\n* **领域特定数据集**：为法律、医疗、技术等专业领域生成数据\n* **数据增强**：通过生成变体来扩充现有的小型数据集\n* **多种格式支持**：支持指令遵循、问答和对话型数据集\n\n\n### **超参数调优模块**\n\n\n\n* 推出 `oumi tune` —— 一个新的超参数搜索与优化模块，用于高效地进行模型调优（#1998、#1991）。感谢 @gbladislau-aumo！\n\n\n### **推理与训练增强**\n\n\n\n* **Bedrock 集成**：新增对 AWS Bedrock 推理引擎的支持，实现可扩展的模型部署（#1983）—— 感谢 @aniruddh-alt！\n* **GKD 训练器支持**：推出新的通用知识蒸馏训练器，适用于模型压缩工作流（#2000）\n* **OpenEnv 强化学习训练**：演示笔记本展示了带有奖励可视化功能的强化学习训练过程（#1996、#2012）\n\n\n### **HPC 与启动器改进**\n\n* **NERSC Perlmutter 支持**：Oumi 启动器现已支持 NERSC Perlmutter HPC 集群（#1959）\n* **增强日志记录**：新增作业日志尾随功能及专用的日志命令，以提升调试效率（#1951、#1964）\n* **惰性云初始化**：优化了启动器的启动性能（#1985）\n\n## **✨ 改进**\n\n\n**模型配置**\n\n* 新增 Qwen3 VL 4B 模型配置（#1992、#1993）\n* 在 ModelParams 中公开 `chat_template_kwargs` 参数，以实现精细化控制（#1997）\n\n**开发者体验**\n\n* 更新 BaseConfig 以支持非原生类型字段（#1684）\n* SLURM 客户端中新增可选的 stdout_file 参数（#1974）\n\n\n## **🐛 错误修复**\n\n\n* 修复了单轮对话数据集中数据分析师出现 NaN 值的问题（#1961）\n* 解决了 SLURM 环境变量相关问题（PMI_RANK → SLURM_PROCID）（#2010）—— 感谢 @AliliRayane！\n* 修复了基础配置中非原生类型字段保存失败的问题（#2005）\n* 更新 uv pip 安装命令，加入 --system 标志（#1979）\n* 通过哈希算法确保推理临时文件名的唯一性（#1986）\n\n\n## **📦 依赖项更新**\n\n\n\n* 升级 transformers：4.56 → 4.57（#1966、#1990）\n* 升级 TRL：0.24.0 → 0.25（#1995、#2011）\n* 锁定 uvicorn 版本以兼容 SkyPilot（#1978）\n\n\n## **🎉 新贡献者**\n\n\n欢迎我们的新贡献者！\n\n\n\n* @gbladislau\n* @oumiandy\n* @AliliRayane\n\n\n## **📖 完整变更日志**\n\n\n如需查看完整的变更列表，请参阅 [完整变更日志](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fcompare\u002Fv0.4.0...v0.5.0)\n","2025-11-18T21:14:34",{"id":163,"version":164,"summary_zh":165,"released_at":166},135931,"v0.4.2","# 发行说明 - v0.4.2\n\n## 🚀 新特性\n\n- **模型支持**：新增对 Qwen3-VL 的支持（[[#1992](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1992)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1992)）\n- **HPC 集群支持**：在 Oumi 启动器中新增对 NERSC Perlmutter HPC 集群的支持（[[#1959](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1959)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1959)）\n- **日志增强**：\n  - 增加了启动器作业的日志跟踪功能（[[#1951](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1951)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1951)）\n  - 新增 `launcher logs` 命令，便于更轻松地访问日志（[[#1964](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1964)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1964)）\n\n## 🐛 错误修复\n\n- 修复了 Sky Pilot 的单元测试（[[#1967](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1967)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1967)）\n- 修复了 GPU 测试中的问题（[[#1970](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1970)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1970)）\n- 锁定 uvicorn 版本以解决与 SkyPilot 的兼容性问题（[[#1978](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1978)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1978)）\n- 更新推理逻辑，始终使用哈希生成唯一的临时文件名（[[#1986](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1986)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1986)）\n- 改进了文档处理相关错误的异常处理机制（[[#1989](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1989)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1989)）\n\n## 🔧 改进\n\n- **性能优化**：在 Oumi 启动器中对云资源进行惰性初始化，以加快启动速度（[[#1985](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1985)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1985)）\n- **代码质量**：\n  - 重构了数据集分析工具（[[#1962](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1962)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1962)，[[#1982](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1982)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1982)）\n  - 将 `conversation_turns` 提取到顶层结构中，以优化数据组织方式（[[#1969](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1969)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1969)）\n  - 使 Slurm 客户端中的 `stdout_file` 参数变为可选（[[#1974](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1974)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1974)）\n\n## 📚 文档更新\n\n- 更新了 README 文件，添加了最新信息（[[#1968](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1968)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1968)）\n- 添加了合成相关的文档及示例配置（[[#1965](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1965)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1965)）\n\n---\n\n**完整变更日志**：https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fcompare\u002Fv0.4.0...v0.4.2","2025-10-20T15:38:02",{"id":168,"version":169,"summary_zh":170,"released_at":171},135932,"v0.4.1","## 变更内容\n* 由 @ryan-arman 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1961 中修复了数据集分析器统计信息中单个对话的 NaN 值。\n* [tiny] 由 @wizeng23 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1963 中为部分测试目录添加了 `__init__.py` 文件。\n* 由 @rlehman221 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1951 中新增了为启动器作业跟踪日志的功能。\n* 由 @ryan-arman 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1962 中将计算统计信息的逻辑移至 `analysis_utils` 模块。\n* 由 @rlehman221 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1967 中修复了 SkyPilot 单元测试失败的问题。\n* 由 @ryan-arman 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1969 中将 `conversation_turns` 从 `conversation_level_summary` 提取到顶层。\n* 由 @wizeng23 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1968 中更新了 `README.md` 文件。\n* [tiny] 由 @wizeng23 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1966 中将 `transformers` 升级至 4.56 版本。\n* 由 @taenin 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1970 中为我们 GPU 测试进行了快速修复。\n* 由 @wizeng23 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1959 中在 Oumi 启动器中支持 NERSC Perlmutter 超算集群。\n* 由 @rlehman221 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1964 中新增了启动器日志命令。\n* 由 @rlehman221 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1974 中使 Slurm 客户端中的 `stdout_file` 参数变为可选。\n* 由 @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1965 中添加了合成文档和示例配置。\n* 由 @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1976 中为调试目的向端到端测试添加了 CUDA 启动阻塞参数。\n* 由 @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1978 中锁定 `uvicorn` 版本以修复 SkyPilot 问题。\n* 由 @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1986 中更新推理逻辑，使其始终使用哈希值生成唯一的推理临时文件名。\n* 由 @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1985 中对 Oumi 启动器中的云资源进行惰性初始化。\n* 由 @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1988 中解除了合成中对 Oumi 数据集支持的阻塞。\n\n\n**完整变更日志**: https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fcompare\u002Fv0.4.0...v0.4.1","2025-10-14T16:35:11",{"id":173,"version":174,"summary_zh":175,"released_at":176},135933,"v0.4.0","# Oumi v0.4 更改日志\n\n## ✨ gpt-oss 训练与推理\n\nOpenAI 于八月发布了备受期待的两款开源权重模型 `gpt-oss-20b` 和 `gpt-oss-120b`。它们是专家混合（MoE）推理模型，具备强大的工具使用能力，并采用原生 4 位量化优化，以实现内存高效的训练和推理。现在您可以在 Oumi 中运行这些模型的训练和推理了！\n\n**使用示例：**\n\n```bash\n# 在单个 GPU 上使用 LoRA 训练 gpt-oss-20b\noumi train -c oumi:\u002F\u002Fconfigs\u002Frecipes\u002Fgpt_oss\u002Fsft\u002F20b_lora_single_gpu_train.yaml\n\n# 使用 vLLM 在本地对 gpt-oss-120b 进行推理\noumi infer -i -c oumi:\u002F\u002Fconfigs\u002Frecipes\u002Fgpt_oss\u002Finference\u002F120b_vllm_infer.yaml\n```\n\n## ⚡ DeepSpeed 支持\n\nDeepSpeed 是一个功能强大且可配置的优化库，支持通过分布式训练和内存优化等技术高效地训练大型模型。Oumi 现在除了 PyTorch 原生的完全分片数据并行（FSDP）之外，也支持 DeepSpeed 进行分布式训练！\n\n**使用示例：**\n\n```bash\n# 使用 DeepSpeed 的 ZeRO-3 优化策略训练 Llama 3.1 8B\noumi train -c oumi:\u002F\u002Fconfigs\u002Fexamples\u002Fdeepspeed\u002Fllama3_1_8b_deepspeed_z3_train.yaml\n\n# 将 DeepSpeed 与 YARN RoPE 缩放结合，以支持更长上下文的训练！\n# 使用 YARN 和 DeepSpeed 训练 Qwen2.5 7B，上下文长度为 128k 个 token\noumi train -c oumi:\u002F\u002Fconfigs\u002Fprojects\u002Flimo\u002Fqwen2.5_7b_fft_yarn_deepspeed.yaml\n```\n\n## 🗄️ Hugging Face 缓存管理 CLI 工具\n\n在使用 Hugging Face Hub 上的数据集和模型时，随着时间的推移，很难追踪哪些内容已被缓存、占用了多少存储空间等信息。在 #1897 中，@aniruddh-alt 向 Oumi CLI 添加了一个 `oumi cache` 工具。该工具允许您查看、添加和删除 Hugging Face Hub 的本地缓存内容，并获取缓存条目的更多信息。\n\n**使用示例：**\n\n```bash\n# 查看缓存中的内容\noumi cache ls\n\n# 过滤包含子字符串“llama”的条目，并按名称排序\noumi cache ls -f *llama* --sort name\n\n# 下载模型到缓存\noumi cache get Qwen\u002FQwen3-0.6B\n\n# 查看缓存模型的信息\noumi cache card Qwen\u002FQwen3-0.6B\n\n# 从缓存中移除模型\noumi cache rm Qwen\u002FQwen3-0.6B\n```\n\n## 🎯 视觉 DPO 和 KTO 支持\n\n我们新增了两种新的训练方法支持：视觉-语言模型上的直接偏好优化（DPO）以及卡尼曼-特沃斯基优化（KTO）。特别感谢 @efsiatras 在 #1538 中实现了 KTO 支持！\n\n**使用示例：**\n\n```bash\n# 对 Qwen2.5-VL 3B 进行视觉 DPO\noumi train -c oumi:\u002F\u002Fconfigs\u002Frecipes\u002Fvision\u002Fqwen2_5_vl_3b\u002Fdpo\u002Ftrain.yaml\n\n# 对 Phi-3 进行 KTO 训练\noumi train -c oumi:\u002F\u002Fconfigs\u002Frecipes\u002Fphi3\u002Fkto\u002Ftrain.yaml\n```\n\n## 🛠️ 开发者体验改进\n\n- 将多个软件包依赖升级至最新版本\n- @penfever 在 #1923 和 #1947 中添加了更多 GGUF、MacOS LlamaCPP 以及远程 Frontier 模型推理配置\n- 在失败时添加预填充的 GitHub 问题链接","2025-09-02T20:38:12",{"id":178,"version":179,"summary_zh":180,"released_at":181},135934,"v0.3.0","# Oumi v0.3 更改日志\n\n## 🔧 模型量化（新功能）\n\n量化是一类至关重要的技术，用于减小模型规模，例如在部署之前。Oumi 现在支持对所有模型应用激活感知权重量化（AWQ）。请参阅我们的 [笔记本](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi%20-%20Quantization%20Tutorial.ipynb) 了解具体操作方法。\n\n**使用示例：**\n\n```python\n# 快速入门 - 将 TinyLlama 量化为 4 位\noumi quantize --method awq_q4_0 --model \"TinyLlama\u002FTinyLlama-1.1B-Chat-v1.0\" --output quantized_model\n\n# 使用配置文件\noumi quantize --config quantization_config.yaml\n```\n\n## ⚖️ 判官 API V2（重大更新）\nLLM-as-a-Judge 是一种利用基础模型来可靠评估其他基础模型的方法。我们已全面重构 Oumi 的 LLM-as-Judge 接口，以提升易用性和灵活性。请查看我们的笔记本 [这里](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi%20-%20Simple%20Judge.ipynb)。\n\n**使用示例：**\n\n```python\nfrom oumi.judges.simple_judge import SimpleJudge\n\n# 内置真实性判官\nsimple_judge = SimpleJudge(judge_config=\"oumi:\u002F\u002Fconfigs\u002Fprojects\u002Fjudges\u002Fgeneric\u002Ftruthfulness.yaml\")\n\ndataset = [{\"request\": \"法国的首都是哪里？\", \"response\": \"罗马\"}]\noutputs = simple_judge.judge(dataset)\n```\n\n## 🎯 自适应推理（新功能）\n\n💪 我们称之为“自适应推理”的功能，是指 Oumi 中新增的特性，可在作业崩溃时恢复训练（或任何任务），以及优化推理并行化以最大化带宽。更多信息请参阅我们的 [笔记本](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi%20-%20Bulk%20Inference%20of%20LLM%20APIs.ipynb)。\n\n## 🛠️ 开发者体验\n\n- 更新了贡献指南\n- 增强了文档\n- 修复了教程笔记本中的问题\n- 改进了错误处理和测试\n- 提升了 MLflow 集成\n- 支持多节点 Slurm 作业\n- 提供了丰富的日志处理器选项\n\n## 新贡献者\n* @amarpal 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1831 中做出了首次贡献\n* @42Shawn 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1837 中做出了首次贡献\n\n**完整更改日志**：https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fcompare\u002Fv0.2.1...v0.3.0","2025-08-05T01:48:50",{"id":183,"version":184,"summary_zh":185,"released_at":186},135935,"v0.2.1","## 变更内容\n* @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1745 中将 infer_online 和 infer_from_file 设置为私有\n* @shanghongsim 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1781 中更新了 launch.md\n* @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1780 中添加了自适应信号量，以支持未来的自适应吞吐量场景\n* @taenin 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1783 中修复了一个 pyright 回归问题\n* @kaisopos 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1782 中对 API V2 进行判断，并修复了从仓库路径加载的判断配置\n* @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1773 中添加了可置换属性和组合采样功能，用于数据合成\n* @shanghongsim 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1788 中移除了微调教程笔记本中的 collator\n* @taenin 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1789 中更新了我们的贡献指南\n* @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1784 中添加了自适应并发控制器，为自适应推理做准备\n* @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1795 中修复了最终对话未能一致保存的问题\n* @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1790 中增加了对合成用数据集摄取的支持\n* @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1791 中增加了对自适应推理的支持\n* @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1797 中为合成添加了示例来源支持\n* @stefanwebb 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1800 中发布了网络研讨会公告及其他新闻\n* @stefanwebb 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1802 中添加了 utm_source 参数\n* @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1796 中添加了处理文档摄取的代码\n* @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1803 中添加了处理基础文档分段的代码\n* @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1804 中更新了 oumi 训练器中的 mflow 支持\n* @wizeng23 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1798 中添加了多节点 verl SLURM 作业\n* @shanghongsim 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1792 中修复了多个教程笔记本\n* @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1807 中为 oumi 训练器添加了参数日志记录功能\n* @kaisopos 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1805 中实现了 API V2 的 YAML 提示变量替换功能\n* @wizeng23 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1809 中对训练配置注释头进行了小幅更新\n* @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1810 中添加了使用 rich 日志处理器的实验性选项\n\n## 新贡献者\n* @shanghongsim 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1781 中做出了首次贡献\n\n**完整变更日志**: https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fcompare\u002Fv0.2.0...v0.2.1","2025-07-11T18:00:36",{"id":188,"version":189,"summary_zh":190,"released_at":191},135936,"v0.2.0","# 亮点\n\n## GRPO 对 trl 和 verl 训练器的支持\n\nOumi 现在同时支持 trl 和 verl 库的 GRPO 训练！这使得您能够使用 Oumi 的配置文件，以零代码或少量代码的方式运行 GRPO 训练。此外，您还可以享受 Oumi 平台的其他功能，例如自定义评估和启动远程任务。\n\n在 Oumi 中运行 GRPO 训练非常简单，只需按照以下步骤操作：\n\n1. 创建一个奖励函数，并使用 `@register(\"\u003Cmy_reward_fn>\", RegistryType.REWARD_FUNCTION)` 将其注册到 Oumi 的奖励函数注册表中。\n2. 创建一个数据集类，将您的 Hugging Face 数据集处理成 [目标框架所需的格式](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fuser_guides\u002Ftrain\u002Ftraining_methods.html#id17)，然后使用 `@register_dataset(\"@hf-org-name\u002Fmy-dataset-name\")` 将其注册到 Oumi 的数据集注册表中。\n3. 使用您的模型、数据集、奖励函数以及超参数创建一个 [Oumi 训练配置](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fuser_guides\u002Ftrain\u002Fconfiguration.html)。有关为 GRPO 设置配置的具体细节，请参阅我们的 [文档](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fuser_guides\u002Ftrain\u002Ftraining_methods.html#group-relative-policy-optimization-grpo)。\n4. 使用 [oumi train CLI](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fuser_guides\u002Ftrain\u002Ftrain.html) 在本地启动训练任务，或者使用 [oumi launch CLI](https:\u002F\u002Foumi.ai\u002Fdocs\u002Fen\u002Flatest\u002Fuser_guides\u002Flaunch\u002Flaunch.html) 启动远程任务。\n\n如需使用 Oumi + trl 的端到端示例，请查看我们的 [笔记本教程](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fnotebooks\u002FOumi%20-%20Train%20a%20Letter%20Counting%20Model%20using%20GRPO.ipynb)。对于 verl，则可以参考我们的多模态 Geometry3K [配置](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fblob\u002Fmain\u002Fconfigs\u002Fexamples\u002Fgrpo_verl_geometry3k\u002Fgcp_job.yaml)。最后，欢迎阅读我们的 [博客文章](https:\u002F\u002Foumi.ai\u002Fblog\u002Fposts\u002Fgrpo-trl-verl)，以获取更多信息。\n\n## 使用 Oumi 构建的模型：HallOumi 和 CoALM\n\n我们很自豪地宣布发布两款由 Oumi 构建的模型：HallOumi 和 CoALM！这两款模型均在 Oumi 上进行训练，我们还提供了从头开始复现其训练过程的配方。\n\n- 🧀 **HallOumi**：一款由 Oumi 开发的真正开源幻觉检测模型，在仅 80 亿参数的情况下，性能超越了 Claude Sonnet、OpenAI o1、DeepSeek R1、Llama 405B 和 Gemini Pro。请在此处查看用于训练该模型的 Oumi 配方：[链接](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Fprojects\u002Fhalloumi)。\n- 🤖 **CoALM**：对话式代理语言模型（CoALM）是一种整合了对话能力和代理能力的统一方法。它包含一个指令微调数据集和三个经过训练的模型（80 亿、700 亿、4050 亿参数）。该项目是伊利诺伊大学香槟分校 ConvAI 实验室与 Oumi 的合作成果，相关论文已被 ACL 接受。请在此处查看用于训练这些模型的 Oumi 配方：[链接](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Ftree\u002Fmain\u002Fconfigs\u002Fprojects\u002Fcoalm)。\n\n## 新增模型支持：Llama 4、Qwen3、Falcon H1 等\n\n我们已新增对众多最新模型的支持。","2025-06-23T19:04:08",{"id":193,"version":194,"summary_zh":195,"released_at":196},135937,"v0.1.14","## 变更内容\n* 在基础推理引擎中记录延迟直方图，由 @nikg4 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1702 中实现\n* 功能：添加 Falcon-E 集成，由 @younesbelkada 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1705 中实现\n* [tiny] 小幅更新以修复未通过的 pre-commit 检查，由 @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1707 中完成\n* 向 DataParams 添加 collator 的 kwargs 字段，由 @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1708 中实现\n* [vision] 添加单独处理图像的选项，由 @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1706 中实现\n* 更新 dev_setup.md 以修正步骤顺序，由 @ryan-arman 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1709 中完成\n* 添加 Molmo 支持的配置，由 @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1710 中实现\n* [tiny] 修复全新安装时的 pre-commit 检查，由 @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1711 中完成\n* 添加 Molmo O 变体的配置，由 @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1712 中实现\n* 添加 Molmo 的实验性 grpo 配置及训练别名，由 @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1713 中实现\n* 更新 installation.md 以修复对 Subversion 的处理，通过添加所需内容……，由 @ryan-arman 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1715 中完成\n* Frontier：修复启动脚本中的 -n 参数，由 @nikg4 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1720 中完成\n* 修复 Falcon H1 的依赖项设置，由 @wizeng23 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1723 中完成\n* 字符计数笔记本的改进，由 @penfever 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1697 中完成\n* [vision] 更新视觉特征生成器，使其支持仅基于补全进行训练，由 @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1722 中实现\n* [tiny] 修复 vl collator 的 bug，由 @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1725 中完成\n* 添加数据合成配置、参数及单元测试，由 @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1700 中实现\n* 为远程推理引擎添加对更多异常类型的支持，并实现对不可重试状态码的快速失败处理，由 @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1704 中完成\n* 添加 Falcon-H1 的 DPO + QLoRA 示例，由 @stefanwebb 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1719 中完成\n* 更新推理流程，使其始终将中间结果写入文件，由 @jgreer013 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1724 中完成\n* 添加关于新 QLoRA 参数的文档，由 @stefanwebb 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1727 中完成\n* Falcon-E 的 README 以及关于所需额外依赖的说明，由 @stefanwebb 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1729 中完成\n* 添加通用视觉数据集，由 @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1726 中实现\n* [tiny][bug] 使 git 命令变为可选，由 @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1730 中完成\n* [tiny][bug] 补充缺失的 Molmo 功能，由 @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1731 中完成\n* [tiny] 更新 phi3-vision 配置以使用 oumi 训练器，由 @oelachqar 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1733 中完成\n* 对启动代码中涉及 2 个云的部分进行小幅错误修复，由 @nikg4 在 https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1728 中完成\n* 上","2025-06-10T20:55:42",{"id":198,"version":199,"summary_zh":200,"released_at":201},135938,"v0.1.13","## What's Changed\r\n* Update dev_setup.md by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1641\r\n* [tiny] Remove vllm install commands by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1643\r\n* Support for custom `processor args`: misc improvements by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1642\r\n* Add Countdown dataset and reward function by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1645\r\n* Adding LoRA train config for Qwen-VL 2.0 by @optas in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1637\r\n* [Evaluation] Convenience function for standard config retrieval by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1644\r\n* Add demo script by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1647\r\n* [bug] fix build errors by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1649\r\n* Adding LoRA train config for SmolVLM by @optas in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1639\r\n* [tiny] Update cli help shorthand by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1648\r\n* Oelachqar\u002Fupdate hooks by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1650\r\n* Add verl PPO trainer by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1646\r\n* Fix a missing dependency in the verl trainer. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1651\r\n* Integrate verl GRPO trainer into train script by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1652\r\n* Update e2e tests to run on lambda by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1653\r\n* Add Qwen3 32B configs by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1661\r\n* Add Qwen3 30B A3B configs by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1665\r\n* [verl] Populate verl config from Oumi config by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1659\r\n* Provide option to configure `label_ignore_index` in training config by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1666\r\n* [Documentation] Custom Evaluations (PR 1-of-2) by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1664\r\n* InterVL-3.0 SFT with *limited* training capabilities by @optas in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1663\r\n* Add verl GRPO Countdown configs by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1668\r\n* Set explicit permissions for our test workflows. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1670\r\n* Add support for repetition_penalty in GrpoParams by @REDDITARUN in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1654\r\n* Fix broken tests due to precommit violations by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1671\r\n* [Documentation] Custom Evaluations (PR 2-of-2) by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1669\r\n* Migrate to `logger.warning` usage by @emmanuel-ferdman in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1673\r\n* Update the Oumi launcher and e2e tests to support runpod. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1672\r\n* Switch back to using GCP for e2e tests. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1675\r\n* Mark an e2e test as is_lora by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1676\r\n* Add Phi4 reasoning plus configs by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1674\r\n* Fix a test breakage caused by a new Click version (8.2.0) by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1679\r\n* chore: edited the link to the stars badge by @Radovenchyk in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1681\r\n* Update verl GRPO countdown configs by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1682\r\n* [very nit] center oumi logo in the cli by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1683\r\n* [tiny] Update training environments doc by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1686\r\n* Add Geometry3K VLM dataset by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1687\r\n* Add `torchao` version to `pyproject.toml` by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1688\r\n* [Feature] Save evaluation config as YAML in output_dir #1546 by @asish-kun in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1680\r\n* Create a script to calculate memory used during training by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1441\r\n* Support VLM-s with VERL_GRPO trainer by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1689\r\n* docs: Add GRPO\u002Fverl documentation by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1690\r\n* Update GRPO letter counting reward function and hparams for stability by @jgreer013 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1692\r\n* [GRPO] Update letter counting notebook by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1694\r\n* Add Lambda Inference Engine by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1695\r\n* Basic shell script for launching jobs on OLCF Frontier HPC cluster by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1691\r\n* Add CoALM dataset class by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1696\r\n* Added exponential backoff and content-type error handling in remote inference engine by @abhiramvad in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1685\r\n* Make SFT datasets usable with GRPO_TRL trainer by @nikg4 in https:\u002F\u002Fgit","2025-05-29T22:12:39",{"id":203,"version":204,"summary_zh":205,"released_at":206},135939,"v0.1.12","## What's Changed\r\n* Add `vllm` to `gpu` optional dependencies by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1614\r\n* [HallOumi] Update inference notebook by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1613\r\n* Update llama4 GCP jobs for non-dev environments. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1621\r\n* Update transformers to 4.51.0 by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1620\r\n* Lazy load skypilot by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1622\r\n* Add additional_model_kwargs and additional_trainer_kwargs to train function by @hommayushi3 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1624\r\n* Added 3 Pixmo vision-language datasets by @jrwana in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1523\r\n* [GRPO] Add notebook to demonstrate GRPO & evaluation for letter counting by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1625\r\n* [Remote Inference] Update Default Params by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1630\r\n* Update trl to 0.16 by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1631\r\n* Support custom `processor args` in `ModelParams` by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1634\r\n* Support BerryBench evaluation by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1635\r\n* [Remote Inference] Error checking for `api_key` by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1638\r\n* Rename cnn_mnist_example to cnn_mnist_tutorial by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1640\r\n* [Remote Inference][GCP] Constructing `api_url` from the Project ID and Region by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1636\r\n\r\n## New Contributors\r\n* @jrwana made their first contribution in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1523\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fcompare\u002Fv0.1.11...v0.1.12","2025-04-16T18:48:38",{"id":208,"version":209,"summary_zh":210,"released_at":211},135940,"v0.1.11","# Oumi v0.1.11 Release Notes 🚀\r\n\r\n## Key Highlights\r\n\r\n### Model Support 🤖\r\n- Integrated Llama 4 (Scout and Maverick variants) with complete workflow configs 🦙\r\n- Added LoRA training for Phi3, Phi4, and Qwen2.5-VL multimodal models 🖼️\r\n\r\n### Developer Experience 💻\r\n- Introduced MLflow integration for experiment tracking 📝\r\n- Enhanced CLI with convenient alias functionality ⌨️\r\n\r\n### HallOumi Framework 🧠\r\n- Added examples for Halloumi\r\n- Added dedicated inference notebooks for both generative and classifier approaches 📓\r\n\r\nWelcome to our new contributors @hommayushi3 and @gabrielaugz! 👋\r\n\r\nFor details, see the [[full changelog](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fcompare\u002Fv0.1.10...v0.1.11)](https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fcompare\u002Fv0.1.10...v0.1.11).\r\n","2025-04-06T22:02:31",{"id":213,"version":214,"summary_zh":215,"released_at":216},135941,"v0.1.10","## What's Changed\r\n* Increment `pillow` version for compatibility with Python 3.13 by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1566\r\n* [Evaluation] Bug: Multiple GPUs attempt to save in the same folder by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1567\r\n\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fcompare\u002Fv0.1.9...v0.1.10","2025-03-25T01:41:04",{"id":218,"version":219,"summary_zh":220,"released_at":221},135942,"v0.1.9","## What's Changed\r\n* Add QwQ full fine-tune and QLoRA configs by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1518\r\n* Update TRL to 0.15 and fix Liger\u002Fdataset code by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1507\r\n* [tiny] Remove vLLM Colab link and fix Alpaca Eval quickstart by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1530\r\n* Evaluation: Inference optimizations by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1522\r\n* Qwen2.5 VL: Replace \"from source\" install with `transformers>=0.49` by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1528\r\n* [Evaluation] Renaming `evaluation_platform` → `evaluation_backend` by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1526\r\n* [tiny] Clean up datasets code by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1529\r\n* Minor logging improvements in `BaseMapDataset` by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1532\r\n* Upload scripts used in a Weekly Walkthrough by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1533\r\n* Update VisionLanguageConversationFeatureGenerator by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1531\r\n* [docs] add security.md by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1534\r\n* [Evaluation] Custom evaluation notebook: a reliability classifier by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1535\r\n* Multimodal: Limit max number of images per Conversation by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1536\r\n* Auto-populate and validate params specific to `vision_language_sft` collator in `TrainingConfig` by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1537\r\n* Update Oumi Env to use Rich formatting by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1541\r\n* Update oumi launch to use Rich formatting by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1543\r\n* Update oumi evaluate to use rich formatting. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1544\r\n* Update the CLI to replace all prints with Rich prints. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1547\r\n* Render the oumi env command as a shell block in bug reports. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1548\r\n* Define `Conversation` proto bufs by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1550\r\n* [Evaluation] Modifying Alpaca Eval results format to be consistent with LM Harness by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1551\r\n* Augmenting logging training\u002Fmodel statistics  by @optas in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1545\r\n* Misc no-op code cleanups by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1553\r\n* Add code used for the evaluation demo. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1556\r\n* Add `OUMI_FORCE_EDITABLE_INSTALL` env var to do editable Oumi install from source in job configs by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1420\r\n* Add letter counting GRPO example by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1539\r\n* Remove UV install from notebooks as this breaks colab by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1558\r\n* [Evaluation] Updates in hallucination notebook by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1552\r\n* [Evaluations] Custom evals: Adding support for `eval_kwargs` by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1557\r\n* Logging message update in `log_number_of_model_parameters` by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1560\r\n* [Evaluation][Custom] Removing restrictions and better error checking by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1561\r\n* Support text truncation (`max_length`) for `vision_language_sft` collator by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1559\r\n* phi 4 multimodal training version 1 ( with limitations ) by @optas in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1555\r\n* Phi-4 basic inference with native\u002Fvllm by @optas in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1563\r\n* [minor] phi4 train improvements by @optas in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1564\r\n* Fix printing errors in oumi env for non-string values. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1565\r\n\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fcompare\u002Fv0.1.8...v0.1.9","2025-03-24T21:27:31",{"id":223,"version":224,"summary_zh":225,"released_at":226},135943,"v0.1.8","## What's Changed\r\n* GRPO trainer: Minimal initial integration by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1482\r\n* Update oumi infer to fall back to interactive mode if no input path is specified. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1483\r\n* Add sample DDP\u002FGCP config for GRPO trainer by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1485\r\n* Temporary fix for chat template issue with multimodal inference w\u002F in-process vLLM engine by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1486\r\n* [tiny] Update async_eval.yaml comments to reference correct class by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1488\r\n* Fix a bug where overriding remote_params fails via the CLI (oumi infer) by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1487\r\n* Define `GrpoParams` under configs by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1490\r\n* Support more GRPO params by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1491\r\n* Minor updates to `oumi env` by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1492\r\n* Warn instead of error when device not found for MFU calculation by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1489\r\n* Updated all CLI endpoints to support oumi:\u002F\u002F prefix by @Spaarsh in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1468\r\n* Fix chat template issue for nested content parts used for VLMs by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1493\r\n* Ctseng777\u002Fjudge by @ctseng777 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1474\r\n* [Evaluation] Modularization & enabling custom evaluations by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1484\r\n* Update documentation formatting for BaseModel by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1494\r\n* Fix `log_samples` not propagating from `eval_kwargs` by @jgreer013 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1496\r\n* [Evaluation] Adding support for logging model samples for all backends by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1499\r\n* Support for deprecated input param (` evaluation_platform`) by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1500\r\n* Limiting the AlpacaEval number of samples for quickstart by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1501\r\n* Add recurring tests to keep our test badges updated. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1498\r\n* Add a schedule for our GPU, CPU, and doc tests by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1503\r\n* Update the GPU Tests badge to use results from main by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1504\r\n* vLLM version increment by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1502\r\n* Minor logging improvements by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1505\r\n* [Evaluation] Save Utils: Moving, fixes, and unit tests by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1506\r\n* Update sample GRPO script to validate num_generations by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1509\r\n* Resolve warning about `--dispatch batches` deprecated param by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1510\r\n* [Evaluation] Re-enabling evaluations with Math Hard (`leaderboard_math_hard`) by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1511\r\n* Update docker image and build script by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1508\r\n* Add Qwen QwQ Lora config by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1514\r\n* Add QwQ eval\u002Finfer configs by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1515\r\n* [Evaluation] Instantiating an inference engine (if needed) when running custom evaluations by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1513\r\n* Switch eval yaml configs to use evaluation_platform by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1516\r\n* Mark `BaseMapDataset` as `typing.Sized` by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1517\r\n* VLM collator refactor by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1512\r\n\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fcompare\u002Fv0.1.7...v0.1.8","2025-03-10T18:25:08",{"id":228,"version":229,"summary_zh":230,"released_at":231},135944,"v0.1.7","## What's Changed\r\n* Update the RemoteInferenceEngine to appropriately handle openai format batch prediction endpoints. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1472\r\n* Fix local models to not break the registry. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1476\r\n* Create an inference config for Claude Sonnet 3.7 by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1479\r\n* Add notebook for fine-tuning MiniMath-R1-1.5B by @jgreer013 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1480\r\n* [Evaluation] Migrate LM Harness integration point from `simple_evaluate` to `evaluate` by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1455\r\n* [tiny]Update trl to 0.14 by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1478\r\n\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fcompare\u002Fv0.1.6...v0.1.7","2025-02-25T23:48:34",{"id":233,"version":234,"summary_zh":235,"released_at":236},135945,"v0.1.6","## What's Changed\r\n* Update RemoteParams to no longer require an API URL. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1452\r\n* [Tiny] Update default training params for Qwen2-VL-2B-Instruct by @optas in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1454\r\n* [Tiny] Add more warnings for \"special\" requirements of Qwen2.5-VL by @optas in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1453\r\n* Minor cleanup of oumi fetch by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1463\r\n* Support for multi-image VLM training by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1448\r\n* Remove a temp workaround in `pad_sequences` on the left side by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1464\r\n* [tiny] Add warning that Oumi doesn't support Intel Macs by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1467\r\n* VLM-related logging improvements  by @nikg4 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1469\r\n* Fix Oumi launcher to be able to run on RunPod and Lambda by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1470\r\n* Enable pre-release install for uv in pyproject.toml by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1466\r\n\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fcompare\u002Fv0.1.5...v0.1.6","2025-02-22T02:25:30",{"id":238,"version":239,"summary_zh":240,"released_at":241},135946,"v0.1.5","## What's Changed\r\n* Fix the remainder of our configs by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1356\r\n* Adopt new Llama 3.1 HF names by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1357\r\n* Define `OUMI_USE_SPOT_VM` env var and start using it to override `use_spot` param  by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1359\r\n* Support HuggingFaceM4\u002FDocmatix dataset by @vishwamartur in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1342\r\n* [nit] update default issue names by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1367\r\n* Update sft_datasets.md by @penfever in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1349\r\n* Have GitHub Trending image hyperlink to GitHub Trending page by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1370\r\n* Update the link for the trending banner. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1371\r\n* Move code to disable caching in `model.config` to a helper function by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1378\r\n* Update transformers version to 4.48 by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1372\r\n* Update notebooks to improve their Colab experience by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1380\r\n* Add proper labels and types to new Bugs and Feature Requests. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1383\r\n* Upgrade omegaconf to 2.4.0dev3 by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1384\r\n* Support HuggingFaceM4\u002Fthe_cauldron dataset by @vishwamartur in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1366\r\n* Update our FAQ for tips about installing oumi on Windows by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1385\r\n* Cleanup `HuggingFaceM4\u002FDocmatix` and `HuggingFaceM4\u002Fthe_cauldron` multimodal datasets by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1387\r\n* Remove uneeded env vars from job configs by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1390\r\n* Remove transformer version override for `HuggingFaceTB\u002FSmolVLM-Instruct` in launcher script by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1388\r\n* [Small Refactor] Moving the inference engine def outside the inference config by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1395\r\n* Evaluation - LM Harness: Adding vLLM support by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1379\r\n* Remove `Docmatix` dataset references from docstrings VLM config examples by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1397\r\n* Fixed broken link in Oumi - A Tour.ipynb notebook by @ciaralema in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1398\r\n* Fix broken links in notebooks. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1402\r\n* Create a client for communicating with a Slurm node via SSH. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1389\r\n* [tiny] Remove references to missing job configs in README by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1404\r\n* Train+Inference with Qwen 2.5 VL (3B) by @optas in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1396\r\n* Add a Slurm cluster and cloud to the oumi launcher. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1406\r\n* Move `pretokenize` script from `scripts\u002Fpretokenize\u002F` to `scripts\u002Fdatasets\u002Fpretokenize\u002F` by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1412\r\n* Create a script to save `Conversation`-s from SFT datasets into  `.jsonl` file by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1413\r\n* [Evaluation] LM Harness refactor by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1410\r\n* Update `save_conversations` tool by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1421\r\n* [SambaNova] Integrate SambaNova Systems to oumi inference by @ctseng777 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1415\r\n* [Μinor] Equating Qwen's 2.5 chat-template to version's 2.0 by @optas in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1419\r\n* Add requirements header to configs and clean them up by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1411\r\n* Updated oumi infer to support CLI argument for system prompt by @Spaarsh in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1422\r\n* [Evaluation] LM Harness remote server support by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1414\r\n* [Feature] Add Tulu3 SFT Mixture Dataset Support by @bwalshe in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1381\r\n* Support Multimodal inference with multiple images and PDF-s in `NATIVE` engine by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1424\r\n* Update notebooks to run on Colab by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1423\r\n* Add calm recipe. by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1425\r\n* Update VLM sample `oumi infer -i` commands by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1428\r\n* Provide example show to start SGLang server using Docker by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1429\r\n* Multi-image support in SGLang inference engine by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1426\r\n* Calm readme by @emrecanacikgoz in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1432\r\n* WildChat-50M Reproduction by @penfever in https:\u002F\u002Fgithub.c","2025-02-20T00:05:43",{"id":243,"version":244,"summary_zh":245,"released_at":246},135947,"v0.1.4","## What's Changed\r\n* Add memory cleanup calls in e2e integration tests  by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1277\r\n* Set up versioning for our documentation by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1275\r\n* Make `qwen2-VL` evaluation job pass by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1278\r\n* Add multi-modal (vlm) notebook with Llama 11B by @optas in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1258\r\n* Documentation: Inference -> List supported models by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1279\r\n* [tiny] update website link by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1280\r\n* Update all documentation links to the new doc URL by @taenin in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1281\r\n* Update Oumi - A Tour.ipynb by @brragorn in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1282\r\n* Documentation: Judge (minor edits) by @kaisopos in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1283\r\n* Fix citation by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1285\r\n* Add Deepseek R1 1.5B\u002F32B configs by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1276\r\n* Misc eval configs cleanup by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1286\r\n* [docs] Describe parallel evaluation by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1284\r\n* Update `microsoft\u002FPhi-3-vision-128k-instruct` training config by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1287\r\n* Add Together Deepseek R1 inference config by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1289\r\n* [minor] vlm notebook minor updates (doc referencing, freeze visual backbone) by @optas in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1288\r\n* Add missing `-m oumi evaluate` argument in eval config by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1291\r\n* [docs] Add more references to VL-SFT and SFT notebooks by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1293\r\n* Eval config change for `deepseek-ai\u002FDeepSeek-R1-Distill-Llama-70B` by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1292\r\n* [notebooks] Update intro & installation instruction by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1294\r\n* Update notebook intros by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1296\r\n* [notebooks] Update installation instructions for colab by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1297\r\n* Add Apache license header to `src\u002Foumi\u002F**\u002F*.py` by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1290\r\n* Minor updates to VLM Multimodal notebook by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1299\r\n* [docs] Add latest notebooks and update references by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1300\r\n* [tiny] Add docs auto-generated `.rst` files to gitignore by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1298\r\n* [tiny] use GitHub link for header by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1301\r\n* [docs][tiny] update inference engines reference by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1302\r\n* Update README\u002Fdocs to add new DeepSeek models by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1304\r\n* [docs] Use `pip install oumi` over `pip install .` by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1305\r\n* Tune VLM SFT configs by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1306\r\n* Tune VLM configs for  SmolVLM and Qwen2-VL by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1307\r\n* Update config\u002Fnotebook pip installs to use PyPI by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1308\r\n* [tiny] upgrade torch version by @oelachqar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1295\r\n* Update logging and unit tests related to chat templates by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1311\r\n* fix(docs): \"interested by joining\" to \"interested in joining\" by @CharlesCNorton in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1312\r\n* Add HF_TOKEN instructions to Oumi Multimodal notebook by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1313\r\n* Update configuration.md by @penfever in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1314\r\n* remove duplicate keys in config example by @lucyknada in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1315\r\n* [Notebooks] Update VLM notebook by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1317\r\n* Update parasail_inference_engine.py by @jgreer013 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1320\r\n* Fix typo and update warning message for OUMI trainer by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1319\r\n* [Notebooks] Add a note that a notebook kernel restart may be needed after `pip install oumi` by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1318\r\n* Update Phi3 to support multiple images by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1321\r\n* Add more detailed comment headers to YAML configs by @wizeng23 in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1310\r\n* [Notebooks] Add a note to Tour notebook to restart kernel after the first `pip install` by @xrdaukar in https:\u002F\u002Fgithub.com\u002Foumi-ai\u002Foumi\u002Fpull\u002F1327\r\n* Tweak `--mem-fraction-static` param in sample SGLang configs by @","2025-02-03T21:06:35"]