search

lxe / simple-llm-finetuner

Simple UI for LLM Model Finetuning
MIT License
2.05k stars 132 forks source link

Not a problem - but like people should know #26

Open Atlas3DSS opened 1 year ago

Atlas3DSS commented 1 year ago

https://arxiv.org/abs/2303.11366 Is a really cool paper about reflection in LLMs image That is after training on like 20 samples for 50 epochs on my 3090 on the 7B model.

User: [Topic or question]

Assistant Hypothetical Response: [Brief or simplified answer to the topic or question]

Agent Reflection: [Critique of the hypothetical response, highlighting the limitations, inaccuracies, or areas that need improvement or expansion, while providing guidance on how to address these issues in the revised response]

Bot Actual Response: [The natural and contextually appropriate answer to the topic or question, as generated by the advanced language model, which incorporates the suggestions and improvements from the agent reflection for a more comprehensive and accurate response]

This + training sets generated with this frame work seem to really improve the generations of these models with fairly limited training sets. Just thought i would share.

lxe commented 1 year ago

Nice to see you can get this from such a small sample set!

Atlas3DSS commented 1 year ago

I have been keeping track of my datasets if anyone else wants to play they are here https://docs.google.com/spreadsheets/d/1QSwJFiyzUQ6H1CloDmJWcHJfYiT7SVxfwBDOOcbvFEo/edit?usp=sharing

Thank you again for making this lovely tool.