recover_model_weight on reward-sim meet problem of _name_or_path and backbone_model_name_or_path about alpaca_farm

[rtaori]

Ok @REIGN12 I believe I have fixed the issue! Please look at the updated instructions at https://github.com/tatsu-lab/alpaca_farm#downloading-pre-tuned-alpacafarm-models.

In particular, first download the sft10k model, and afterwards downloading the reward models (to the same parent directory) should work.

[REIGN12]

I think for people easy to use your reward model, the following hack may save the day:

1. change the config.json in hf repo: change the backbone_name_or_path to any public llama weight(diff) huggingface repo
2. change the logic in common.py the function get_pretrained_model_name_with_model_name_or_path, to better handle this logic

Just some of my thoughts! Hope you guys may fix this bug soon! ❤️

[rtaori]

Ah sorry about this issue! Taking a look into it now