Ray2333/GRM-Llama3.2-3B-rewardmodel-ft

Introduction

This reward model achieves a score of 90.9 on reward-bench, which is finetuned from the Ray2333/GRM-llama3.2-3B-sftreg using the decontaminated Skywork preference dataset v0.2. We obtain a SOTA 3B reward model that can outperform a series of 8B reward models and even surpass gpt4/gemini as a judge.

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

Evaluation

We evaluate GRM-Llama3.2-3B-rewardmodel-ft on the reward model benchmark, where it achieved strong performance among models smaller than 7B.

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

Model	Average	Chat	Chat Hard	Safety	Reasoning
GRM-Llama3-8B-rewardmodel-ft(8B)	91.5	95.5	86.2	90.8	93.6
GRM-Llama3.2-3B-rewardmodel-ft(ours, 3B)	90.9	91.6	84.9	92.7	94.6
GRM-gemma2-2B-rewardmodel-ft (Ours, 2B)	88.4	93.0	77.2	92.2	91.2
google/gemini-1.5-pro-0514	88.2	92.3	80.6	87.9	92.0
RLHFlow/pair-preference-model-LLaMA3-8B	87.1	98.3	65.8	89.7	94.7
GRM-llama3-8B-sftreg(ours, 8B)	87.0	98.6	67.8	89.2	92.3
google/gemini-1.5-pro-0924	86.8	94.1	77.0	85.8	90.2
openai/gpt-4o-2024-08-06	86.7	96.1	76.1	88.1	86.6
GRM-llama3.2-3B-sftreg(ours, 3B)	85.8	96.4	67.1	88.2	91.6
GRM-Gemma-2B-rewardmodel-ft (Ours, 2B)	84.7	89.4	75.2	85.5	88.8
openai/gpt-4o-2024-05-13	84.6	96.6	70.4	86.5	84.9
sfairXC/FsfairX-LLaMA3-RM-v0.1 (8B)	84.4	99.4	65.1	86.8	86.4
Nexusflow/Starling-RM-34B	82.6	96.9	57.2	87.7	88.5
GRM-Gemma2-2B-sftreg(Ours, 2B)	81.0	97.2	59.6	86.9	80.3
GRM-Gemma-2B-sftreg(Ours, 2B)	75.3	95.5	48.7	80.0	76.8
berkeley-nest/Starling-RM-7B-alpha (7B)	74.6	98	43.4	88.6	74.6
Gemma-2B-rewardmodel-baseline(Ours, 2B)	73.7	94.1	46.1	79.6	75.0
openbmb/UltraRM-13b (13B)	71.3	96.1	55.3	45.8	82

Usage

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

device = 'cuda:0'
# load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained('Ray2333/GRM-Llama3.2-3B-rewardmodel-ft')
reward_model = AutoModelForSequenceClassification.from_pretrained(
                'Ray2333/GRM-Llama3.2-3B-rewardmodel-ft', torch_dtype=torch.float16, 
                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: "<bos><start_of_turn>user\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?<end_of_turn>\n<start_of_turn>model\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?<end_of_turn>\n".

kwargs = {"padding": 'max_length', "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))[0]
  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}
}