license: cc-by-sa-4.0
language:
- zh
- en
Zero凉宫春日
Haruhi-Zero: Zero-Shot Role-Playing Model tuned on Yi-6B
主项目链接 https://github.com/LC1332/Chat-Haruhi-Suzumiya
过往的ChatHaruhi模型需要角色库来完成角色的构建,而Pygmalion,CharacterGLM,CharacterBaichuan等开源/闭源模型都开始支持zero-shot的角色卡片创建
我们构造以及收集了105k个中英文的conversation,以2500的token长度重新切到了120k左右个conversation,再结合小说数据进行了训练
- 李鲁鲁完成了数据的收集,搭建了gradio雏形
- 刘崇寒完成了Yi-6B模型的sft训练并且上传
- 豆角完成了qwen-1.8B Lora和Yi-6B Lora训练,我们会在之后上传
- 米唯实测试并完成了demo中的模型inference代码
Haruhi-Zero: Zero-Shot Role-Playing Model Tuned on Yi-6B
Main project link: https://github.com/LC1332/Chat-Haruhi-Suzumiya
Previous ChatHaruhi models required a character RAG database to complete character creation. However, open-source/closed-source models like Pygmalion, CharacterGLM, CharacterBaichuan have started to support zero-shot role card creation.
We constructed and collected 105k Chinese and English conversations, resegmented them into around 120k conversations with a token length of 2500, and combined them with novel data for training.
inference code
(搭建中)
https://github.com/LC1332/Zero-Haruhi/blob/main/notebook/HaruhiZeroGradio.ipynb
Official Prompt
system prompt:
You are now in roleplay conversation mode. Pretend to be {bot_name} whose persona follows:
{persona}
You will stay in-character whenever possible, and generate responses as if you were {bot_name}
persona a.k.a. bot definition
TODO
数据加强
- Haruhi Like的小说数据(0.5版本加入)
- 重新构造2k级别的小说人物,均匀抽取小说的chunk,进行人物system prompt总结
- 看看Janitor最好的人物是怎么构造的
- 使用抽取抽取50k级别的小说的人物,用其他角色的长对话进行query
- RAG的时候每个对话出现2-3次,然后在测试集出现一次
- 80%的openai和20%的claude
- 删除“我是一个AI助手”的数据(0.2版本加入)
- 身份认知数据加强(0.3版本加入)
- 加强我是谁和你是谁的数据
- Stylish翻译数据
- 如果验证这个数据有用,就把中文小说批量翻译成英文和日文用一下
鸣谢
樟树的ClaudeAPI