Fine-tuning framework?
Hi,
Which framework that you for fine-tuning? I was trying to adapt ToolBench (https://github.com/OpenBMB/ToolBench), a function calling dataset to the Yi series,
But the source code works for LLaMA2 but not for Yi seris, the train loss was 0.0 and eval loss being NaN.
I am looking for a working/stable QLoRA framework for open source LLMs where users simply need to bring their models and curated datasets.
Thanks!
I use an a version of the original qlora training script that I've adapted over time; it's rather hacked together at this point, but it works. This model, however, uses Yi-34b as adapted to the LLama2 architecture (https://huggingface.co/chargoddard/Yi-34B-Llama), so in principle you should be able to get ToolBench to work if it does for other llama2 models (though I have no experience with ToolBench). Note that I use the model from the llama-tokenizer branch of that model repo to also remove any dependency on the Yi tokenizer definition. I haven't tried training native Yi with the custom model and/or tokenizer definitions.