search

strickvl / mlops-dot-systems

Quarto technical blog
4 stars 0 forks source link

posts/2024-06-15-isafpr-first-finetune #8

Open utterances-bot opened 2 weeks ago

utterances-bot commented 2 weeks ago

Alex Strick van Linschoten - Finetuning my first LLM(s) for structured data extraction with axolotl

I finetuned my first LLM(s) for the task of extracting structured data from ISAF press releases. Initial tests suggest that it worked pretty well out of the box.

https://mlops.systems/posts/2024-06-15-isafpr-first-finetune.html

saeedesmaili commented 2 weeks ago

Thanks for the write up. Interestingly, I struggle with Mistral as my finetuned model rarely outputs the token and it continues generating tokens, but I don't have this issue with finetuned llama3-8b models. I couldn't figure out what I'm doing wrong with Mistral that leads to this issue.

The template-free data format is interesting, but I can't understand what does it mean to mask some parts of the training data so model wouldn't learn them. My intuition is we want our fine-tuned model to learn everything in the training dataset (including the system prompt, the instructions, and the response). I need to dig into the axolotl docs to read more on this.