silk-road/Haruhi-Baize-Role-Playing-Conversation
收藏资源简介:
--- license: cc-by-4.0 task_categories: - text-generation language: - zh --- ## Haruhi-Zero的Conversation训练数据 我们计划拓展ChatHaruhi,从Few-shot到Zero-shot,这个数据集记录使用各个(中文)角色扮演api进行Baize式相互聊天后得到的数据结果 ids代表聊天的时候两张bot的角色卡片, 角色卡片的信息可以在https://huggingface.co/datasets/silk-road/Haruhi-Zero-RolePlaying-movie-PIPPA 中找到 并且对于第一次出现的id0,也会在prompt字段中进行记录。 聊天的时候id和ids的卡片进行对应 - openai 代表两个聊天的bot都使用openai - GLM 代表两个聊天的bot都使用CharacterGLM - Claude 代表两个聊天的bot都使用Claude - Claude_openai 代表id0的使用Claude, id1的使用openai - Baichuan 代表两个聊天的bot都使用Character-Baichuan-Turbo 目前百川有很严重的访问限制,如果谁有并发更大的百川的接口,可以联系我们借用一下(邮箱chengli.thu@gmail.com) 或者跑下面的代码,(start_id = 10590 end_id = 12708)反馈给我 https://github.com/LC1332/Zero-Haruhi/blob/main/notebook/GenerateBaizeBaichuan.ipynb 目前平衡生成时间、成本和效果来看,最终训练准备先采用openai和Claude_openai, 这两者已经采集了15000/2000的数据,正在进一步生成更多数据 主项目链接 https://github.com/LC1332/Chat-Haruhi-Suzumiya ## API和服务器 如果你有OpenAI、Claude或者Character-Baichuan的api资源 可以参与进来的话,方便联系我一下 发邮件或者在知乎 https://www.zhihu.com/people/cheng-li-47 留一下您的微信 如果你有足够的训练资源去tuning 13B以及Yi 34B规模的模型(2000长度),也可以联系我加入到项目后面的训练中。 如果你能够组织Human Feedback的比较标注,也可以联系我们 整体计划 https://o9z6tor1qu.feishu.cn/docx/LxTWdGnP2oQ0oUx8H0wcmyZCnrb 后期如果有空的话可以进一步拿各家api的结果做一下RLHF或者DPO。
license: cc-by-4.0 task_categories: - text-generation language: - zh ## Haruhi-Zero's Conversation Training Data We plan to expand ChatHaruhi from Few-shot to Zero-shot scenarios. This dataset records the outcomes of mutual chat sessions conducted in a Baize-style manner between two bots using various Chinese large language model APIs for role-playing. The `ids` field denotes the role cards of the two bots during the chat. Detailed information of the role cards can be accessed at https://huggingface.co/datasets/silk-road/Haruhi-Zero-RolePlaying-movie-PIPPA. Moreover, the first occurrence of id0 will also be documented in the `prompt` field. The id and its corresponding role card are matched during the chat. ### API Category Definitions - `openai`: Both chatting bots utilize OpenAI models - `GLM`: Both chatting bots utilize CharacterGLM - `Claude`: Both chatting bots utilize Claude models - `Claude_openai`: id0 uses Claude, while id1 uses OpenAI - `Baichuan`: Both chatting bots utilize Character-Baichuan-Turbo Currently, Baichuan API has extremely strict access restrictions. If you have Baichuan API interfaces with higher concurrency limits, please contact us for borrowing (email: chengli.thu@gmail.com). Alternatively, you can run the following code snippet (start_id = 10590, end_id = 12708) and share the results with us. The corresponding Jupyter notebook is available at https://github.com/LC1332/Zero-Haruhi/blob/main/notebook/GenerateBaizeBaichuan.ipynb Considering the balance of generation time, cost and performance, we initially intend to use OpenAI and Claude_openai datasets for training. 15,000 out of 20,000 samples for these two categories have been collected, and we are currently generating additional data. Main project repository: https://github.com/LC1332/Chat-Haruhi-Suzumiya --- ## API and Server Resources If you have access to OpenAI, Claude or Character-Baichuan APIs and wish to participate in the project, please contact us via email or leave your WeChat ID on Zhihu at https://www.zhihu.com/people/cheng-li-47. If you possess sufficient training resources to fine-tune 13B and Yi 34B scale models (with a sequence length of 2000 tokens), you are welcome to join us for the subsequent training phase. If you are able to organize human feedback annotation work, please also contact us. Overall project plan: https://o9z6tor1qu.feishu.cn/docx/LxTWdGnP2oQ0oUx8H0wcmyZCnrb If time permits in the future, we will further conduct RLHF or DPO training using the results from various API models.
Haruhi-Zero的Conversation训练数据
数据集描述
- 任务类别: 文本生成
- 语言: 中文
数据集内容
- 角色卡片: 数据集记录使用各个(中文)角色扮演API进行Baize式相互聊天后得到的数据结果。角色卡片的信息可以在silk-road/Haruhi-Zero-RolePlaying-movie-PIPPA中找到。
- ID和IDS: ids代表聊天时两张bot的角色卡片,对于第一次出现的id0,会在prompt字段中进行记录。
使用的API
- openai: 两个聊天的bot都使用openai
- GLM: 两个聊天的bot都使用CharacterGLM
- Claude: 两个聊天的bot都使用Claude
- Claude_openai: id0使用Claude,id1使用openai
- Baichuan: 两个聊天的bot都使用Character-Baichuan-Turbo
数据生成
- 当前状态: 目前平衡生成时间、成本和效果来看,最终训练准备先采用openai和Claude_openai,这两者已经采集了15000/2000的数据,正在进一步生成更多数据。
联系方式
- 邮箱: chengli.thu@gmail.com
- 代码链接: GenerateBaizeBaichuan.ipynb
- 主项目链接: Chat-Haruhi-Suzumiya




