You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi, the model is trained with Llama 1 and we release the weight differences (the delta of weights) instead of the actual weights. To recover the weights for training and inference, first, you need to download the Llama 1 base model (e.g. https://huggingface.co/luodian/llama-7b-hf), then run the script apply_delta to recover the model. Explained arguments are as follows.
--base-model-path (the Llama 1 base model you need to download and save at this dir)
--target-model-path (the output dir for the recovered model)
--delta-path (the character-ll-wdiff dir, which you can download from our repo)
PS: We use Llama 1 because, at that time, there were not many choices of good open-source LLMs for fine-tuning. You can always switch to a more powerful LLM (e.g. Llama 3) to train a better character-llm.
I'm a computer nerd, but very interested in big language modelling
`python3 -m fastchat.model.apply_delta \
character-llm-beethoven-7b-wdiff ----> the model that can be downloaded here (there's a big pytorch_model file)
And where do the models llama-7b and character-llm-beethoven-7b correspond to the content?
Can you be more specific?
The text was updated successfully, but these errors were encountered: