Add support for pre-trained reward models

[diegofiori]

Description

OpenAssistant has released on HF the reward models they trained on the open-source datasets. Even if they are not tailored for the user need, we could lavarege them as a starting point for fine-tuning the user reward models.

Available reward models:

• OpenAssistant/reward-model-deberta-v3-large-v2 · Hugging Face
• OpenAssistant/reward-model-deberta-v3-large · Hugging Face
• OpenAssistant/reward-model-deberta-v3-base · Hugging Face
• OpenAssistant/reward-model-electra-large-discriminator · Hugging Face

TODO

• [ ] Add the possibility to load the pretrained checkpoints for the reward model in ChatLLaMA
• [ ] Write unittests

[gagan3012]

Can I work on this?

[PierpaoloSorbellini]

Please go ahead, let me know if you need any support or if you have any questions. I assigned you to this issue. Thank you! @gagan3012