[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"tool-microsoft--GODEL":3,"similar-microsoft--GODEL":131},{"id":4,"github_repo":5,"name":6,"description_en":7,"description_zh":8,"ai_summary_zh":8,"readme_en":9,"readme_zh":10,"quickstart_zh":11,"use_case_zh":12,"hero_image_url":13,"owner_login":14,"owner_name":15,"owner_avatar_url":16,"owner_bio":17,"owner_company":18,"owner_location":18,"owner_email":19,"owner_twitter":20,"owner_website":21,"owner_url":22,"languages":23,"stars":52,"forks":53,"last_commit_at":54,"license":55,"difficulty_score":56,"env_os":57,"env_gpu":58,"env_ram":59,"env_deps":60,"category_tags":71,"github_topics":75,"view_count":90,"oss_zip_url":18,"oss_zip_packed_at":18,"status":91,"created_at":92,"updated_at":93,"faqs":94,"releases":130},10167,"microsoft\u002FGODEL","GODEL","Large-scale pretrained models for goal-directed dialog","GODEL 是一款由微软研发的大规模预训练模型，专为“目标导向型对话”设计。与传统闲聊机器人不同，它擅长处理需要达成具体任务或依赖外部信息的复杂对话场景，例如根据检索到的文档回答问题、执行日程提醒或调用 API 完成特定指令。\n\nGODEL 有效解决了传统对话模型在面对需要结合上下文与外部知识（如数据库、搜索结果）时表现不佳的难题。通过在海量的多轮对话数据及指令数据上进行预训练，它能够精准地将外部文本信息融入回复生成过程中，显著提升了在零样本（zero-shot）及少样本场景下的任务完成度。\n\n该工具主要面向 AI 研究人员和开发者。基于 Transformer 编码器 - 解码器架构，GODEL 提供了灵活的微调机制，用户仅需少量特定任务的数据，即可将其快速适配到客服助手、任务型智能体等新场景中。项目基于 Huggingface Transformers 构建，提供了完整的源代码、数据集及预训练模型权重，支持便捷的本地部署与二次开发，是构建下一代实用型对话系统的有力基石。","# GODEL: Large-Scale Pre-Training for Goal-Directed Dialog\n\n## News\n\n(Update 10\u002F23\u002F2022) We have released GODEL V1.1, which is trained on 551M multi-turn dialogs from Reddit discussion thread, and 5M instruction and knowledge grounded dialogs. It has shown significantly better results on our benchmark, especially in the zero-shot setting.\n\nPlease check out our model cards in the huggingface Transformers repository. With several lines of code, it should be pretty straightforward to chat with GODEL. A live demo is shown [here.](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fmicrosoft\u002FGODEL-Demo)\n\nBase model: https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FGODEL-v1_1-base-seq2seq\n\nLarge model: https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FGODEL-v1_1-large-seq2seq\n\n## Introduction\nThis repository showcases **building goal-directed dialog** using GODEL, and contains the dataset, source code and pre-trained model for the following paper:\n\n\n[GODEL: Large-Scale Pre-Training for Goal-Directed Dialog](https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fpublication\u002Fgodel-large-scale-pre-training-for-goal-directed-dialog\u002F)\u003Cbr>Baolin Peng, Michel Galley, Pengcheng He, Chris Brockett, Lars Liden, Elnaz Nouri, Zhou Yu, Bill Dolan, Jianfeng Gao\n![image](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmicrosoft_GODEL_readme_2528af09c5ad.png)\n\nGODEL is a large-scale pre-trained model for goal-directed dialogs. It is parameterized with a Transformer-based encoder-decoder model and trained for response generation grounded in external text, which allows more effective fine-tuning on dialog tasks that require conditioning the response on information that is external to the current conversation (e.g., a retrieved document). The pre-trained model can be efficiently fine-tuned and adapted to accomplish a new dialog task with a handful of task-specific dialogs.\n\nThis repository is based on Hugginface Transformers. Some evaluation scripts and dataset are adapted from [DSTC7-End-to-End-Conversation-Modeling](data\u002Fgrounded), [DialoGPT](data\u002Fungrounded), [UnifiedQA](https:\u002F\u002Fgithub.com\u002Fallenai\u002Funifiedqa), [MS MARCO](https:\u002F\u002Fmicrosoft.github.io\u002Fmsmarco\u002F), [MultiWOZ](https:\u002F\u002Fgithub.com\u002Fbudzianowski\u002Fmultiwoz), [Schema-Guided Dataset](https:\u002F\u002Fgithub.com\u002Fgoogle-research-datasets\u002Fdstc8-schema-guided-dialogue), etc.\n\nThe included scripts can be used to reproduce the results reported in the paper. Project and demo webpage: [https:\u002F\u002Faka.ms\u002FGODEL](https:\u002F\u002Faka.ms\u002FGODEL)\n\n## Installation \n**Requires** The interactive interface requries *node.js* and *npm*. Please refer to [here](https:\u002F\u002Fdocs.npmjs.com\u002Fdownloading-and-installing-node-js-and-npm) for installation.\n\nPlease use the below commands to create the environment, clone the repo and install required packages.\n```\nconda create -n godel-env python=3.8\nconda activate godel-env\nconda install nodejs\ngit clone https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FGODEL.git\ncd GODEL\npip install -r requirements.txt\nexport PYTHONPATH=\"`pwd`\"\n```\nFetch and unzip the pretrained model based on which to continue finetune your own data.  \n\n```zsh\nwget https:\u002F\u002Fbapengstorage.blob.core.windows.net\u002Ffileshare\u002Fgodel_base.tar.gz\ntar -zxvf godel_base.tar.gz\n```\n## Pipeline\n**Data format**\n```json\n  {\n    \"Context\": \"Please remind me of calling to Jessie at 2PM.\",\n    \"Knowledge\": \"reminder_contact_name is Jessie, reminder_time is 2PM\",\n    \"Response\": \"Sure, set the reminder: call to Jesse at 2PM\"\n  },\n```\nWe use json format to represent a training example. As shown in the above example, it contains the following fields:\n* **Context** - The context from session beginning to current turn.\n* **Knowledge** - External or environment state represented in plain text.\n* **Reponse** - The target agent respose. It can be a template, an api call or natural language.\n\n**Fine-tuning**\n```Bash\nDATA_NAME={path_of_data}\nOUTPUT_DIR={path_of_fine-tuned_model}\nMODEL_PATH={path_of_pre-trained_model}\nEXP_NAME={experiment_name}\n\npython train.py --model_name_or_path ${MODEL_PATH} \\\n\t--dataset_name ${DATA_NAME} \\\n\t--output_dir ${OUTPUT_DIR} \\\n\t--per_device_train_batch_size=16 \\\n\t--per_device_eval_batch_size=16 \\\n\t--max_target_length 512 \\\n\t--max_length 512 \\\n\t--num_train_epochs 50 \\\n\t--save_steps 10000 \\\n\t--num_beams 5 \\\n\t--exp_name ${EXP_NAME} --preprocessing_num_workers 24\n```\n\n\n**Generation**\n```python\nDATA_NAME={path_of_data}\nOUTPUT_DIR={path_to_save_predictions}\nMODEL_PATH={path_of_fine-tuned_model}\n\npython generate.py --model_name_or_path ${MODEL_PATH}  \\\n\t--dataset_name ${DATA_NAME}  \\\n\t--output_dir ${OUTPUT_DIR}  \\\n\t--per_device_eval_batch_size=16  \\\n\t--max_target_length 128 \\\n\t--max_length 512  \\\n\t--preprocessing_num_workers 24  \\\n\t--num_beams 5 \n```\n\n**Interaction**  \n\nWe provide a demo interface to chat with finetuned models. The backend server is based on *flask* and the interface is based on *vue*, *bootstrap-vue*, and *BasicVueChat*.\n\nStart the backend server:\n```bash\n# Please create the backend server refering to e.g., dstc9_server.py\npython EXAMPLE_server.py # start the sever and expose 8080 \n```\n\nStart serving frontend page:\n```bash\ncd GODEL\u002Fhtml\nnpm install\nnpm run serve \n```\nOpen localhost:8080 in your web browser, you will see the following page. Note that the backend port should be consistent with the port used in html\u002Fcompoents\u002Fchat.vue.\n\nA live demo is shown [here.](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fmicrosoft\u002FGODEL-Demo)\n\n## Models\n\nWe have released GODEL V1.1, which is trained on 551M multi-turn dialogs from Reddit discussion thread and 5M instruction and knowledge-grounded dialogs. More models will be released later.\n\n~~We have released three fine-tuned models which can be further fine-tuned on low-resource user-customized dataset. The total parameters in these models range from 117M to 2.7B.~~\n\n| Model      | Huggingface Model Cards |\n| :---: | :---: |\n| Base      |  [microsoft\u002FGODEL-v1_1-base-seq2seq](https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FGODEL-v1_1-base-seq2seq)      |\n| Large   |     [microsoft\u002FGODEL-v1_1-large-seq2seq](https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FGODEL-v1_1-large-seq2seq)    |\n\n\n## Training\n\n5\u002F22\u002F2023: Pretraining GODEL models with our codebase is no longer supported, but GODEL models remain available. See [here](TRAIN.md) for details.\n\n### Fine-tuning and Evaluation\n\nGODEL is fine-tuned and evaluated on four tasks. We provide scripts to create training and testing data in our format. Please refer to *create_downstream_dataset.sh* to download the original data and execute the following cmd.\n\n```Bash\ncd scripts \n.\u002Fcreate_downstream_dataset.sh\n```\n\n```Bash\nGROUNDED_CHECKPOINT={path_to_saved_checkpoint}\nOUTPUT_DIR={path_to_save_predictions}\nTASK=wow\naccelerate launch --config_file configs\u002FG16_config.yaml train.py \n\t--model_name_or_path ${GROUNDED_CHECKPOINT} \\\n\t--dataset_name .\u002Fdatasets_loader\u002F${TASK}_dataset.py \\\n\t--output_dir ${OUTPUT_DIR} \\\n\t--per_device_train_batch_size=16 \\\n\t--per_device_eval_batch_size=16 \\\n\t--max_target_length 256 \\\n\t--max_length 512 \\\n\t--num_train_epochs 10 \\\n\t--preprocessing_num_workers 24 \\\n\t--num_beams 5 \\\n\t--exp_name ${TASK}  \\\n\t--learning_rate 5e-5 \\\t\n\t--save_every_checkpoint \\\n\t--save_steps 50000 \n```\n\n## Tutorial - Adding a new task using GODEL\n\nIn this tutorial, you will build a grounded dialog model based on GODEL for DSTC9 task. Detailed information can be found at [here](https:\u002F\u002Fgithub.com\u002Falexa\u002Falexa-with-dstc9-track1-dataset).\n\nFirstly download the data and convert it to GODEL format.\n```bash\ncd examples\u002Fdstc9\n.\u002Fcreate_data.sh\n```\n*Finetune with the pre-trained GODEL model*\n```bash\ncd GODEL \nGODEL_MODEL={path_to_pre-trained_model}\npython train.py \n\t--model_name_or_path ${GODEL_MODEL}   \\\n\t--dataset_name ..\u002Fexamples\u002Fdstc9\u002Fdstc9_dataset.py   \\\n\t--output_dir ..\u002Fexamples\u002Fdstc9\u002Fckpt   \\\n\t--per_device_train_batch_size=16  \\\n\t--per_device_eval_batch_size=16  \\\n\t--max_target_length 128  \\\n\t--max_length 512  \\\n\t--num_train_epochs 50  \\\n\t--save_steps 10000  \\\n\t--num_beams 5  \\\n\t--exp_name wow-test \\\n\t--preprocessing_num_workers 24 \\\n\t--save_every_checkpoint \n```\n*Interact with above trained model*\n```bash\ncd examples\u002Fdstc9\n# replace model path in dstc9_server with a trained ckpt in line 49\npython dstc9_server.py\n\ncd GODEL\u002Fhtml \nnpm install\nnpm run serve\n```\n\n## Disclaimer\nThis repository aims to facilitate research in a paradigm shift of building task bots at scale. This toolkit contains only part of the modeling machinery needed to actually produce a model weight file in a running dialog. On its own, this model provides only information about the weights of various text spans; in order for a researcher to actually use it, they will need to bring in-house conversational data of their own for future pre-training and decode the response generation from the pretrained\u002Ffinetuned system. Microsoft is not responsible for any generation from the 3rd party utilization of the pretrained system.\n\n\u003C!-- ## Contact\nShould you have any questions\u002Fsuggestions, feel free to contact bapeng@microsoft.com. -->\n\n## Citation\nif you use this code and data in your research, please cite our arxiv paper:\n```\n@misc{peng2022godel,\nauthor = {Peng, Baolin and Galley, Michel and He, Pengcheng and Brockett, Chris and Liden, Lars and Nouri, Elnaz and Yu, Zhou and Dolan, Bill and Gao, Jianfeng},\ntitle = {GODEL: Large-Scale Pre-training for Goal-Directed Dialog},\nhowpublished = {arXiv},\nyear = {2022},\nmonth = {June},\nurl = {https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fpublication\u002Fgodel-large-scale-pre-training-for-goal-directed-dialog\u002F},\n}\n```\n\n\n\u003C!-- # Project\n\n> This repo has been populated by an initial template to help get you started. Please\n> make sure to update the content to build a great experience for community-building.\n\nAs the maintainer of this project, please make a few updates:\n\n- Improving this README.MD file to provide a great experience\n- Updating SUPPORT.MD with content about this project's support experience\n- Understanding the security reporting process in SECURITY.MD\n- Remove this section from the README -->\n\n## Contributing\n\nThis project welcomes contributions and suggestions.  Most contributions require you to agree to a\nContributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us\nthe rights to use your contribution. For details, visit https:\u002F\u002Fcla.opensource.microsoft.com.\n\nWhen you submit a pull request, a CLA bot will automatically determine whether you need to provide\na CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions\nprovided by the bot. You will only need to do this once across all repos using our CLA.\n\nThis project has adopted the [Microsoft Open Source Code of Conduct](https:\u002F\u002Fopensource.microsoft.com\u002Fcodeofconduct\u002F).\nFor more information see the [Code of Conduct FAQ](https:\u002F\u002Fopensource.microsoft.com\u002Fcodeofconduct\u002Ffaq\u002F) or\ncontact [opencode@microsoft.com](mailto:opencode@microsoft.com) with any additional questions or comments.\n\n## Trademarks\n\nThis project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft \ntrademarks or logos is subject to and must follow \n[Microsoft's Trademark & Brand Guidelines](https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Flegal\u002Fintellectualproperty\u002Ftrademarks\u002Fusage\u002Fgeneral).\nUse of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship.\nAny use of third-party trademarks or logos are subject to those third-party's policies.\n","# GODEL：面向目标导向对话的大规模预训练\n\n## 新闻\n\n（更新日期：2022年10月23日）我们发布了GODEL V1.1，该模型基于Reddit讨论串中的5.51亿轮多轮对话，以及500万条指令和知识增强型对话进行训练。在我们的基准测试中，它表现出了显著更好的效果，尤其是在零样本设置下。\n\n请查看Hugging Face Transformers仓库中的模型卡片。只需几行代码，就可以轻松地与GODEL进行对话。实时演示请见[这里](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fmicrosoft\u002FGODEL-Demo)。\n\n基础模型：https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FGODEL-v1_1-base-seq2seq\n\n大型模型：https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FGODEL-v1_1-large-seq2seq\n\n## 简介\n本仓库展示了如何使用GODEL构建目标导向对话，并包含了以下论文所需的数据集、源代码和预训练模型：\n\n\n[GODEL：面向目标导向对话的大规模预训练](https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fpublication\u002Fgodel-large-scale-pre-training-for-goal-directed-dialog\u002F)\u003Cbr>彭宝林、米歇尔·加利、何鹏程、克里斯·布罗克特、拉斯·利登、埃尔纳兹·努里、周宇、比尔·多兰、高剑峰\n![image](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmicrosoft_GODEL_readme_2528af09c5ad.png)\n\nGODEL是一个用于目标导向对话的大规模预训练模型。它采用基于Transformer的编码器-解码器架构，并在外部文本基础上进行响应生成训练，这使得在需要根据当前对话之外的信息（例如检索到的文档）来调整响应的对话任务上，能够更有效地进行微调。该预训练模型可以通过少量特定任务的对话数据进行高效微调，从而适应新的对话任务。\n\n本仓库基于Hugging Face Transformers框架。部分评估脚本和数据集改编自[DSTC7-端到端对话建模](data\u002Fgrounded)、[DialoGPT](data\u002Fungrounded)、[UnifiedQA](https:\u002F\u002Fgithub.com\u002Fallenai\u002Funifiedqa)、[MS MARCO](https:\u002F\u002Fmicrosoft.github.io\u002Fmsmarco\u002F)、[MultiWOZ](https:\u002F\u002Fgithub.com\u002Fbudzianowski\u002Fmultiwoz)、[Schema-Guided Dataset](https:\u002F\u002Fgithub.com\u002Fgoogle-research-datasets\u002Fdstc8-schema-guided-dialogue)等。\n\n包含的脚本可用于复现论文中报告的结果。项目及演示网页：[https:\u002F\u002Faka.ms\u002FGODEL](https:\u002F\u002Faka.ms\u002FGODEL)\n\n## 安装 \n**要求** 交互式界面需要*node.js*和*npm*。请参考[此处](https:\u002F\u002Fdocs.npmjs.com\u002Fdownloading-and-installing-node-js-and-npm)进行安装。\n\n请使用以下命令创建环境、克隆仓库并安装所需包。\n```\nconda create -n godel-env python=3.8\nconda activate godel-env\nconda install nodejs\ngit clone https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FGODEL.git\ncd GODEL\npip install -r requirements.txt\nexport PYTHONPATH=\"`pwd`\"\n```\n获取并解压预训练模型，以便继续用您自己的数据进行微调。  \n\n```zsh\nwget https:\u002F\u002Fbapengstorage.blob.core.windows.net\u002Ffileshare\u002Fgodel_base.tar.gz\ntar -zxvf godel_base.tar.gz\n```\n## 流程\n**数据格式**\n```json\n  {\n    \"Context\": \"请提醒我下午2点给杰西打电话。\",\n    \"Knowledge\": \"reminder_contact_name 是杰西，reminder_time 是下午2点\",\n    \"Response\": \"好的，已设置提醒：下午2点给杰西打电话\"\n  },\n```\n我们使用JSON格式表示一个训练样本。如上例所示，它包含以下字段：\n* **Context** - 从会话开始到当前轮次的上下文。\n* **Knowledge** - 以纯文本形式表示的外部或环境状态。\n* **Response** - 目标代理的响应。它可以是模板、API调用或自然语言。\n\n**微调**\n```Bash\nDATA_NAME={数据路径}\nOUTPUT_DIR={微调后模型保存路径}\nMODEL_PATH={预训练模型路径}\nEXP_NAME={实验名称}\n\npython train.py --model_name_or_path ${MODEL_PATH} \\\n\t--dataset_name ${DATA_NAME} \\\n\t--output_dir ${OUTPUT_DIR} \\\n\t--per_device_train_batch_size=16 \\\n\t--per_device_eval_batch_size=16 \\\n\t--max_target_length 512 \\\n\t--max_length 512 \\\n\t--num_train_epochs 50 \\\n\t--save_steps 10000 \\\n\t--num_beams 5 \\\n\t--exp_name ${EXP_NAME} --preprocessing_num_workers 24\n```\n\n\n**生成**\n```python\nDATA_NAME={数据路径}\nOUTPUT_DIR={保存预测结果的路径}\nMODEL_PATH={微调后模型路径}\n\npython generate.py --model_name_or_path ${MODEL_PATH}  \\\n\t--dataset_name ${DATA_NAME}  \\\n\t--output_dir ${OUTPUT_DIR}  \\\n\t--per_device_eval_batch_size=16  \\\n\t--max_target_length 128 \\\n\t--max_length 512  \\\n\t--preprocessing_num_workers 24  \\\n\t--num_beams 5 \n```\n\n**交互**  \n\n我们提供了一个演示界面，用于与微调后的模型进行对话。后端服务器基于*flask*，界面则基于*vue*、*bootstrap-vue*和*BasicVueChat*。\n\n启动后端服务器：\n```bash\n# 请参照例如dstc9_server.py创建后端服务器\npython EXAMPLE_server.py # 启动服务器并暴露8080端口\n```\n\n启动前端页面服务：\n```bash\ncd GODEL\u002Fhtml\nnpm install\nnpm run serve \n```\n在您的浏览器中打开localhost:8080，您将看到如下页面。请注意，后端端口应与html\u002Fcompoents\u002Fchat.vue中使用的端口一致。\n\n实时演示请见[这里](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fmicrosoft\u002FGODEL-Demo)。\n\n## 模型\n\n我们发布了GODEL V1.1，该模型基于Reddit讨论串中的5.51亿轮多轮对话，以及500万条指令和知识增强型对话进行训练。后续还将发布更多模型。\n\n~~我们已经发布了三款微调模型，这些模型可以进一步在低资源的用户自定义数据集上进行微调。这些模型的总参数量范围为1.17亿至27亿。~~\n\n| 模型      | Hugging Face 模型卡片 |\n| :---: | :---: |\n| 基础版      |  [microsoft\u002FGODEL-v1_1-base-seq2seq](https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FGODEL-v1_1-base-seq2seq)      |\n| 大型版   |     [microsoft\u002FGODEL-v1_1-large-seq2seq](https:\u002F\u002Fhuggingface.co\u002Fmicrosoft\u002FGODEL-v1_1-large-seq2seq)    |\n\n\n## 训练\n\n2023年5月22日：使用我们的代码库对GODEL模型进行预训练已不再支持，但GODEL模型仍然可用。详情请参阅[此处](TRAIN.md)。\n\n### 微调与评估\n\nGODEL 在四个任务上进行了微调和评估。我们提供了以我们的格式创建训练和测试数据的脚本。请参考 *create_downstream_dataset.sh* 来下载原始数据，并执行以下命令。\n\n```Bash\ncd scripts \n.\u002Fcreate_downstream_dataset.sh\n```\n\n```Bash\nGROUNDED_CHECKPOINT={保存检查点的路径}\nOUTPUT_DIR={保存预测结果的路径}\nTASK=wow\naccelerate launch --config_file configs\u002FG16_config.yaml train.py \n\t--model_name_or_path ${GROUNDED_CHECKPOINT} \\\n\t--dataset_name .\u002Fdatasets_loader\u002F${TASK}_dataset.py \\\n\t--output_dir ${OUTPUT_DIR} \\\n\t--per_device_train_batch_size=16 \\\n\t--per_device_eval_batch_size=16 \\\n\t--max_target_length 256 \\\n\t--max_length 512 \\\n\t--num_train_epochs 10 \\\n\t--preprocessing_num_workers 24 \\\n\t--num_beams 5 \\\n\t--exp_name ${TASK}  \\\n\t--learning_rate 5e-5 \\\t\n\t--save_every_checkpoint \\\n\t--save_steps 50000 \n```\n\n## 教程 - 使用 GODEL 添加新任务\n\n在本教程中，您将基于 GODEL 构建一个用于 DSTC9 任务的 grounded 对话模型。详细信息请参见 [这里](https:\u002F\u002Fgithub.com\u002Falexa\u002Falexa-with-dstc9-track1-dataset)。\n\n首先下载数据并将其转换为 GODEL 格式。\n```bash\ncd examples\u002Fdstc9\n.\u002Fcreate_data.sh\n```\n*使用预训练的 GODEL 模型进行微调*\n```bash\ncd GODEL \nGODEL_MODEL={预训练模型的路径}\npython train.py \n\t--model_name_or_path ${GODEL_MODEL}   \\\n\t--dataset_name ..\u002Fexamples\u002Fdstc9\u002Fdstc9_dataset.py   \\\n\t--output_dir ..\u002Fexamples\u002Fdstc9\u002Fckpt   \\\n\t--per_device_train_batch_size=16  \\\n\t--per_device_eval_batch_size=16  \\\n\t--max_target_length 128  \\\n\t--max_length 512  \\\n\t--num_train_epochs 50  \\\n\t--save_steps 10000  \\\n\t--num_beams 5  \\\n\t--exp_name wow-test \\\n\t--preprocessing_num_workers 24 \\\n\t--save_every_checkpoint \n```\n*与上述训练好的模型交互*\n```bash\ncd examples\u002Fdstc9\n# 将 dstc9_server 中第 49 行的模型路径替换为已训练的检查点\npython dstc9_server.py\n\ncd GODEL\u002Fhtml \nnpm install\nnpm run serve\n```\n\n## 免责声明\n本仓库旨在促进大规模构建任务型机器人这一范式转变的研究。该工具包仅包含实际生成运行对话所需模型权重文件的部分建模机制。单独使用此模型只能提供关于各个文本片段权重的信息；研究人员若要真正使用它，还需引入自己的对话数据进行后续预训练，并从预训练或微调后的系统中解码响应生成过程。微软不对第三方利用预训练系统所产生的任何内容负责。\n\n\u003C!-- ## 联系方式\n如有任何疑问或建议，请随时联系 bapeng@microsoft.com。 -->\n\n## 引用\n如果您在研究中使用了本代码和数据，请引用我们的 arXiv 论文：\n```\n@misc{peng2022godel,\nauthor = {Peng, Baolin and Galley, Michel and He, Pengcheng and Brockett, Chris and Liden, Lars and Nouri, Elnaz and Yu, Zhou and Dolan, Bill and Gao, Jianfeng},\ntitle = {GODEL: Large-Scale Pre-training for Goal-Directed Dialog},\nhowpublished = {arXiv},\nyear = {2022},\nmonth = {June},\nurl = {https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Fresearch\u002Fpublication\u002Fgodel-large-scale-pre-training-for-goal-directed-dialog\u002F},\n}\n```\n\n\n\u003C!-- # 项目\n\n> 此仓库已填充初始模板，以帮助您快速入门。请务必更新内容，以打造良好的社区建设体验。\n\n作为该项目的维护者，请进行以下几项更新：\n\n- 改进 README.MD 文件，以提供良好的用户体验\n- 更新 SUPPORT.MD，添加关于该项目支持体验的内容\n- 理解 SECURITY.MD 中的安全报告流程\n- 从 README 中移除本节 -->\n\n## 贡献\n本项目欢迎贡献和建议。大多数贡献都需要您同意贡献者许可协议（CLA），声明您有权且确实授予我们使用您贡献的权利。有关详情，请访问 https:\u002F\u002Fcla.opensource.microsoft.com。\n\n当您提交拉取请求时，CLA 机器人会自动判断您是否需要提供 CLA，并相应地标记 PR（例如状态检查、评论）。只需按照机器人提供的指示操作即可。对于所有使用我们 CLA 的仓库，您只需完成一次此步骤。\n\n本项目已采用 [微软开源行为准则](https:\u002F\u002Fopensource.microsoft.com\u002Fcodeofconduct\u002F)。更多信息请参阅 [行为准则常见问题解答](https:\u002F\u002Fopensource.microsoft.com\u002Fcodeofconduct\u002Ffaq\u002F)，或发送电子邮件至 [opencode@microsoft.com](mailto:opencode@microsoft.com) 咨询更多问题或意见。\n\n## 商标\n本项目可能包含项目、产品或服务的商标或标识。未经授权使用微软商标或标识须遵守并遵循 [微软商标与品牌指南](https:\u002F\u002Fwww.microsoft.com\u002Fen-us\u002Flegal\u002Fintellectualproperty\u002Ftrademarks\u002Fusage\u002Fgeneral)。在本项目的修改版本中使用微软商标或标识不得造成混淆或暗示微软的赞助。任何第三方商标或标识的使用均受其各自政策的约束。","# GODEL 快速上手指南\n\nGODEL 是一个用于**目标导向对话（Goal-Directed Dialog）**的大规模预训练模型。它基于 Transformer Encoder-Decoder 架构，擅长结合外部知识（如检索文档）生成回复，适用于任务型机器人开发。\n\n## 1. 环境准备\n\n在开始之前，请确保您的系统满足以下要求：\n\n*   **操作系统**: Linux 或 macOS (Windows 需使用 WSL)\n*   **Python**: 3.8 及以上版本\n*   **包管理器**: Conda (推荐)\n*   **前端依赖**: Node.js 和 npm (仅当需要运行本地交互界面时需要)\n*   **硬件**: 推荐使用 NVIDIA GPU 进行微调或推理\n\n## 2. 安装步骤\n\n### 2.1 创建虚拟环境并安装依赖\n\n使用以下命令创建名为 `godel-env` 的 Conda 环境，克隆代码库并安装 Python 依赖：\n\n```bash\nconda create -n godel-env python=3.8\nconda activate godel-env\nconda install nodejs\ngit clone https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FGODEL.git\ncd GODEL\npip install -r requirements.txt\nexport PYTHONPATH=\"`pwd`\"\n```\n\n> **提示**: 如果下载速度较慢，可配置国内 pip 源（如清华源）：\n> `pip install -r requirements.txt -i https:\u002F\u002Fpypi.tuna.tsinghua.edu.cn\u002Fsimple`\n\n### 2.2 获取预训练模型\n\n下载并解压基础预训练模型（Base Model），以便进行后续的微调或推理：\n\n```zsh\nwget https:\u002F\u002Fbapengstorage.blob.core.windows.net\u002Ffileshare\u002Fgodel_base.tar.gz\ntar -zxvf godel_base.tar.gz\n```\n\n*注：您也可以直接从 Hugging Face 加载模型，无需手动下载文件（见下文“基本使用”）。*\n\n## 3. 基本使用\n\nGODEL 提供了多种使用方式，以下是两种最常用的场景。\n\n### 场景一：直接使用 Hugging Face Transformers 进行对话（最简单）\n\n如果您只想快速体验或与模型交互，无需运行完整的仓库代码，可直接使用 `transformers` 库。\n\n**前提**: 确保已安装 `transformers` (`pip install transformers`)\n\n```python\nfrom transformers import AutoTokenizer, AutoModelForSeq2SeqLM\n\n# 加载分词器和模型 (Base 版本)\nmodel_name = \"microsoft\u002FGODEL-v1_1-base-seq2seq\"\ntokenizer = AutoTokenizer.from_pretrained(model_name)\nmodel = AutoModelForSeq2SeqLM.from_pretrained(model_name)\n\n# 准备输入数据\n# Context: 对话上下文\n# Knowledge: 外部知识\u002F信息\ncontext = \"Please remind me of calling to Jessie at 2PM.\"\nknowledge = \"reminder_contact_name is Jessie, reminder_time is 2PM\"\n\n# GODEL 的输入格式通常为：Context + Knowledge\ninput_text = f\"{context} [KNOWLEDGE] {knowledge}\"\n\n# 编码\ninputs = tokenizer(input_text, return_tensors=\"pt\")\n\n# 生成回复\noutputs = model.generate(\n    **inputs,\n    max_length=128,\n    num_beams=5,\n    early_stopping=True\n)\n\n# 解码结果\nresponse = tokenizer.decode(outputs[0], skip_special_tokens=True)\nprint(f\"AI Response: {response}\")\n```\n\n### 场景二：微调到自定义任务\n\n若需将 GODEL 适配到特定领域，需准备 JSON 格式数据并运行微调脚本。\n\n**1. 数据格式示例 (`data.json`)**\n```json\n{\n  \"Context\": \"Please remind me of calling to Jessie at 2PM.\",\n  \"Knowledge\": \"reminder_contact_name is Jessie, reminder_time is 2PM\",\n  \"Response\": \"Sure, set the reminder: call to Jesse at 2PM\"\n}\n```\n\n**2. 执行微调命令**\n替换路径变量后运行：\n\n```Bash\nDATA_NAME=.\u002Fpath\u002Fto\u002Fyour\u002Fdata.json\nOUTPUT_DIR=.\u002Foutput_finetuned_model\nMODEL_PATH=.\u002Fgodel_base  # 或者 HuggingFace 模型路径\n\npython train.py --model_name_or_path ${MODEL_PATH} \\\n\t--dataset_name ${DATA_NAME} \\\n\t--output_dir ${OUTPUT_DIR} \\\n\t--per_device_train_batch_size=16 \\\n\t--per_device_eval_batch_size=16 \\\n\t--max_target_length 512 \\\n\t--max_length 512 \\\n\t--num_train_epochs 50 \\\n\t--save_steps 10000 \\\n\t--num_beams 5 \\\n\t--exp_name my_task --preprocessing_num_workers 24\n```\n\n### 场景三：启动本地交互演示 (Web UI)\n\n如果需要图形化界面与微调后的模型聊天：\n\n1.  **启动后端服务** (修改 `EXAMPLE_server.py` 中的模型路径):\n    ```bash\n    python EXAMPLE_server.py\n    ```\n2.  **启动前端页面**:\n    ```bash\n    cd GODEL\u002Fhtml\n    npm install\n    npm run serve \n    ```\n3.  **访问**: 打开浏览器访问 `http:\u002F\u002Flocalhost:8080`。\n\n---\n*更多详情及在线 Demo 体验，请访问 [Hugging Face Spaces](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fmicrosoft\u002FGODEL-Demo)。*","某电商平台的智能客服团队正在升级其售后系统，旨在让机器人能直接读取订单数据库和物流文档，精准处理用户的复杂退换货请求。\n\n### 没有 GODEL 时\n- **信息割裂严重**：传统对话模型无法有效结合外部知识库，面对用户“我的订单 #12345 为什么还没发货”的提问，只能回复通用套话，无法调取实时物流状态。\n- **开发成本高昂**：每新增一个业务场景（如退款、换货），都需要人工编写大量规则或收集数千条特定语料进行从头训练，迭代周期长达数周。\n- **多轮对话易迷失**：在长流程交互中，模型容易遗忘用户之前提供的关键信息（如收货地址、商品瑕疵描述），导致反复询问用户，体验极差。\n\n### 使用 GODEL 后\n- **知识精准落地**：GODEL 原生支持基于外部文本生成回复，能直接将检索到的订单详情和物流文档融入对话，准确回答“您的包裹因暴雨滞留中转站，预计明天发出”。\n- **小样本快速适配**：凭借大规模预训练能力，仅需几十条针对新业务的示范对话，GODEL 即可通过微调掌握新任务，将新场景上线时间缩短至几天内。\n- **上下文记忆稳固**：其编码器 - 解码器架构专为多轮目标导向对话设计，能全程追踪用户意图与关键实体，无需重复确认信息即可流畅完成复杂的售后闭环。\n\nGODEL 通过将外部知识与对话逻辑深度对齐，让智能客服从“只会聊天的机器人”进化为“能办实事的业务专家”。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmicrosoft_GODEL_e1e69e39.png","microsoft","Microsoft","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002Fmicrosoft_4900709c.png","Open source projects and samples from Microsoft",null,"opensource@microsoft.com","OpenAtMicrosoft","https:\u002F\u002Fopensource.microsoft.com","https:\u002F\u002Fgithub.com\u002Fmicrosoft",[24,28,32,36,40,44,48],{"name":25,"color":26,"percentage":27},"Python","#3572A5",87.6,{"name":29,"color":30,"percentage":31},"Makefile","#427819",4.3,{"name":33,"color":34,"percentage":35},"SCSS","#c6538c",4.2,{"name":37,"color":38,"percentage":39},"Vue","#41b883",1.9,{"name":41,"color":42,"percentage":43},"Shell","#89e051",1.5,{"name":45,"color":46,"percentage":47},"HTML","#e34c26",0.4,{"name":49,"color":50,"percentage":51},"JavaScript","#f1e05a",0.1,886,113,"2026-04-07T20:18:17","MIT",3,"Linux, macOS, Windows","未说明 (基于 Transformer 架构，通常建议 NVIDIA GPU 进行训练\u002F推理，具体显存需求取决于模型大小)","未说明 (建议 16GB+ 以支持数据处理和多进程预处理)",{"notes":61,"python":62,"dependencies":63},"1. 交互式界面需要安装 Node.js 和 npm。\n2. 后端服务器基于 Flask，前端基于 Vue。\n3. 需手动下载预训练模型文件（如 godel_base.tar.gz）。\n4. 自 2023 年 5 月 22 日起，不再支持使用该代码库进行预训练，仅支持微调和评估。\n5. 建议使用 conda 创建名为 'godel-env' 的环境进行管理。","3.8",[64,65,66,67,68,69,70],"torch","transformers","accelerate","nodejs","npm","flask","vue",[72,73,74],"语言模型","数据工具","开发框架",[76,77,78,79,80,81,65,82,83,84,85,86,87,88,89],"data-processing","dialogue","dialogue-systems","machine-learning","text-data","text-generation","conversational-ai","language-grounding","grounded-generation","dialogpt","language-model","pretrained-model","pytorch","transformer",2,"ready","2026-03-27T02:49:30.150509","2026-04-20T20:44:28.085170",[95,100,105,110,115,120,125],{"id":96,"question_zh":97,"answer_zh":98,"source_url":99},45638,"如何复现论文中的评估指标（如 F1^R, F1^K）？","官方离线计算相关指标（如 F1, KF1），使用的是 ParlAI 的官方实现。您可以参考以下链接获取具体的度量标准代码：https:\u002F\u002Fparl.ai\u002Fdocs\u002Ftutorial_metrics.html#list-of-metrics。此外，注意 `generate.py` 中存在小 bug，需将第 478 行的 `eval_dataloader` 改为 `dataloader`，第 534 行的 `eval_dataloader` 改为 `test_dataloader`。","https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FGODEL\u002Fissues\u002F16",{"id":101,"question_zh":102,"answer_zh":103,"source_url":104},45639,"运行 train.py 时出现 KeyError: 'test' 错误怎么办？","该问题是因为脚本默认缺少测试文件参数。已提交的 PR #18 修改了 `train.py` 以支持 `--test_file` 参数。使用自定义数据集时，请参考以下命令格式运行：\n`python train.py --model_name_or_path ..\u002Fmodels\u002FGODEL-Base\u002F --train_file [训练集路径] --validation_file [验证集路径] --test_file [测试集路径] --output_dir [输出目录] --exp_name godel2 --num_train_epochs 10 --per_device_train_batch_size 16 --per_device_eval_batch_size 16 --max_target_length 32`\n请确保您的代码已更新到包含此修复的版本。","https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FGODEL\u002Fissues\u002F17",{"id":106,"question_zh":107,"answer_zh":108,"source_url":109},45640,"在哪里可以找到示例中引用的 `DialoGLM.server` 模块？","示例代码中引用的 `from DialoGLM.server import *` 路径可能有误。正确的导入方式应直接指向项目主目录下的 server 模块，即使用：`from GODEL.server import *` 或直接导入 `GODEL\u002Fserver.py` 文件。","https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FGODEL\u002Fissues\u002F13",{"id":111,"question_zh":112,"answer_zh":113,"source_url":114},45641,"是否会公开带有用户人工评估分数的生成数据集？","由于法律方面的顾虑（legal concerns），维护团队目前没有计划公开包含人工评估分数的数据集。","https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FGODEL\u002Fissues\u002F14",{"id":116,"question_zh":117,"answer_zh":118,"source_url":119},45642,"requirements.txt 中存在两个冲突的 nltk 版本，应该使用哪一个？","`requirements.txt` 文件中同时列出了 `nltk==3.7` 和 `nltk==3.4`，这会导致依赖冲突。通常建议保留较新的稳定版本（即 `nltk==3.7`），并删除旧版本条目以解决安装报错。","https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FGODEL\u002Fissues\u002F35",{"id":121,"question_zh":122,"answer_zh":123,"source_url":124},45643,"模型微调后的输出文件中为什么全是 \"todo\"？","如果微调后的模型输出（如 `output\u002Ftest-step-NNN`）全为 `[\"todo\"]`，这通常不是模型本身的问题，而是由测试数据格式不正确或内容异常导致的。请仔细检查您的测试数据是否符合 GODEL 格式要求，并确保数据预处理无误。","https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FGODEL\u002Fissues\u002F23",{"id":126,"question_zh":127,"answer_zh":128,"source_url":129},45644,"如何使用特定的解码策略和 Prompt 模板来复现结果？","论文中提到使用了 beam search（束搜索，beam=5）。关于 Prompt 模板，训练数据中知识接地（knowledge grounded）的对话通常包含 `\u003C|knowledge|>` token。虽然社区用户在尝试复现时遇到了一些困难（如生成随机数据或重复知识），但建议参考训练数据格式：上下文与知识之间使用特定 token 分隔。若遇到生成 `unk |knowledge|>` 的情况，请检查输入数据是否严格遵循了包含 `Context`, `Knowledge`, `Response` 字段的 JSON 格式，并确认是否正确加载了指令微调（Instruction Tuning）相关的配置。","https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FGODEL\u002Fissues\u002F25",[],[132,142,151,159,167,175],{"id":133,"name":134,"github_repo":135,"description_zh":136,"stars":137,"difficulty_score":56,"last_commit_at":138,"category_tags":139,"status":91},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,"2026-04-06T06:32:30",[140,74,141,73],"Agent","图像",{"id":143,"name":144,"github_repo":145,"description_zh":146,"stars":147,"difficulty_score":90,"last_commit_at":148,"category_tags":149,"status":91},9989,"n8n","n8n-io\u002Fn8n","n8n 是一款面向技术团队的公平代码（fair-code）工作流自动化平台，旨在让用户在享受低代码快速构建便利的同时，保留编写自定义代码的灵活性。它主要解决了传统自动化工具要么过于封闭难以扩展、要么完全依赖手写代码效率低下的痛点，帮助用户轻松连接 400 多种应用与服务，实现复杂业务流程的自动化。\n\nn8n 特别适合开发者、工程师以及具备一定技术背景的业务人员使用。其核心亮点在于“按需编码”：既可以通过直观的可视化界面拖拽节点搭建流程，也能随时插入 JavaScript 或 Python 代码、调用 npm 包来处理复杂逻辑。此外，n8n 原生集成了基于 LangChain 的 AI 能力，支持用户利用自有数据和模型构建智能体工作流。在部署方面，n8n 提供极高的自由度，支持完全自托管以保障数据隐私和控制权，也提供云端服务选项。凭借活跃的社区生态和数百个现成模板，n8n 让构建强大且可控的自动化系统变得简单高效。",184740,"2026-04-19T23:22:26",[73,74,140,141,150],"插件",{"id":152,"name":153,"github_repo":154,"description_zh":155,"stars":156,"difficulty_score":56,"last_commit_at":157,"category_tags":158,"status":91},10095,"AutoGPT","Significant-Gravitas\u002FAutoGPT","AutoGPT 是一个旨在让每个人都能轻松使用和构建 AI 的强大平台，核心功能是帮助用户创建、部署和管理能够自动执行复杂任务的连续型 AI 智能体。它解决了传统 AI 应用中需要频繁人工干预、难以自动化长流程工作的痛点，让用户只需设定目标，AI 即可自主规划步骤、调用工具并持续运行直至完成任务。\n\n无论是开发者、研究人员，还是希望提升工作效率的普通用户，都能从 AutoGPT 中受益。开发者可利用其低代码界面快速定制专属智能体；研究人员能基于开源架构探索多智能体协作机制；而非技术背景用户也可直接选用预置的智能体模板，立即投入实际工作场景。\n\nAutoGPT 的技术亮点在于其模块化“积木式”工作流设计——用户通过连接功能块即可构建复杂逻辑，每个块负责单一动作，灵活且易于调试。同时，平台支持本地自托管与云端部署两种模式，兼顾数据隐私与使用便捷性。配合完善的文档和一键安装脚本，即使是初次接触的用户也能在几分钟内启动自己的第一个 AI 智能体。AutoGPT 正致力于降低 AI 应用门槛，让人人都能成为 AI 的创造者与受益者。",183572,"2026-04-20T04:47:55",[140,72,150,74,141],{"id":160,"name":161,"github_repo":162,"description_zh":163,"stars":164,"difficulty_score":56,"last_commit_at":165,"category_tags":166,"status":91},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",[74,141,140],{"id":168,"name":169,"github_repo":170,"description_zh":171,"stars":172,"difficulty_score":90,"last_commit_at":173,"category_tags":174,"status":91},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",161692,"2026-04-20T11:33:57",[74,140,72],{"id":176,"name":177,"github_repo":178,"description_zh":179,"stars":180,"difficulty_score":90,"last_commit_at":181,"category_tags":182,"status":91},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。",109154,"2026-04-18T11:18:24",[74,141,140]]