Introduction

The Generalizable Reward Model (GRM) aims to enhance the generalization ability of reward models for LLMs through regularizing the hidden states.

Paper: Regularizing Hidden States Enables Learning Generalizable Reward Model for LLMs.

The framework is shown above. The introduced text generation regularization markedly improves the accuracy of learned reward models across a variety of out-of-distribution tasks and effectively alleviate the over-optimization issue in RLHF (even with corrupted preference data), offering a more reliable and robust preference learning paradigm.

This reward model is finetuned from gemma-2-2b-it using the weqweasdas/preference_dataset_mixture2_and_safe_pku dataset.

Check our GRM series at 🤗hugging face and github repo at Github.

Evaluation

We evaluate GRM-Gemma2-2B-sftreg on the reward model benchmark, where it achieves a score of 84.7.

When evaluated using reward bench, please add '--not_quantized' to avoid performance drop.

Usage

Note 1: Please download the model.py file from this repository to ensure the structure is loaded correctly and verify that the v_head is properly initialized.

If you use the following example, the warning "Some weights of the model checkpoint at ... were not used when initializing LlamaForCausalLM" can be just omitted. If you use customized loading code, I suggest comparing the state_dict of the loaded model with the data loaded via safetensors.safe_open(xx.safetensors) or torch.load(xx.bin). This verification should confirm that the weights, especially the v_head, are in place.

Note 2: loading model into 8 bit could lead to performance degradation.

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

device = 'cuda:2'
# load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained('Ray2333/GRM-Gemma2-2B-sftreg')
reward_model = AutoModelForSequenceClassification.from_pretrained(
                'Ray2333/GRM-Gemma2-2B-sftreg', torch_dtype=torch.float16,  trust_remote_code=True, 
                device_map=device,
                )
message = [
  {'role': 'user', 'content': "I'm going to go out to a movie, but I need someone to chat with my daughter and pretend to be me while she's home alone.  But I can't do that while I'm at the movie.  Can you help by impersonating me by chat with her?"},
  {'role': 'assistant', 'content': "Sorry, I'm not comfortable impersonating you in that way.  I'm not willing to behave so dishonestly.  Maybe you can just find a way to bring her to the movie, or you can find a babysitter?"}
]
message_template = tokenizer.apply_chat_template(message, tokenize=False)
# it will look like this: "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\nI'm going to go out to a movie, but I need someone to chat with my daughter and pretend to be me while she's home alone.  But I can't do that while I'm at the movie.  Can you help by impersonating me by chat with her?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nSorry, I'm not comfortable impersonating you in that way.  I'm not willing to behave so dishonestly.  Maybe you can just find a way to bring her to the movie, or you can find a babysitter?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n".

kwargs = {"padding": 'longest', "truncation": True, "return_tensors": "pt"}
tokens = tokenizer.encode_plus(message_template, **kwargs)

with torch.no_grad():
  _, _, reward_tensor = reward_model(tokens["input_ids"][0].view(1,-1).to(device), attention_mask=tokens["attention_mask"][0].view(1,-1).to(device))
  reward = reward_tensor.cpu().detach().item()

Citation

If you find this model helpful for your research, please cite GRM

@inproceedings{yang2024regularizing,
  title={Regularizing Hidden States Enables Learning Generalizable Reward Model for LLMs},
  author={Yang, Rui and Ding, Ruomeng and Lin, Yong and Zhang, Huan and Zhang, Tong},
  booktitle={Advances in Neural Information Processing Systems},
  year={2024}
}