44670
44670
是4096，但是vllm等推理框架默认会限制最大长度，需要修改配置。
`System: `这个是不需要的。并且前面的那些文本也可以删除。 模型没有特殊的提示标题，直接把系统提示放在最前面，并要求模型始终按照系统提示回答即可。加上`System: `可能会有反效果。
可以试一下基于mistral2-7b-v20.2进一步微调 也有可能是数据集太少导致的
thank you for your suggestion!
Thank you for your diligent work on the Falcon model. By the way, we are currently working on training a new version of the 40B Falcon model, which will have...
非常感谢！期待能早些见到我们的模型上线~ 我们最近训练了7B参数的v1.4版本，已经上传到了hf：https://huggingface.co/OpenBuddy/openbuddy-llama-7b-v1.4-fp16 7B v1.4在处理各类任务上均有较强的提升，在我们内部的测试里优于13B模型。此外，7B模型对显卡的需求较低，4090应当可用，推荐你们选择这个版本。 附我们在使用的System Prompt，供参考： ``` Consider a conversation between User (a human) and Assistant (named Buddy). Buddy is an INTP-T, a friendly, intelligent and multilingual AI assistant, by...
Refractoring the current api design might be needed for adding password support.
I have found an issue might be related: https://github.com/LostRuins/koboldcpp/issues/1026
According to https://github.com/ollama/ollama/issues/6048#issuecomment-2258749861, this issue should have been resolved in 0.3.1.
I have also met this issue with bnb+olora