Quantile Regression for Distributional Reward Models in RLHF

Author: Nicolai Dorka
Tech Report: https://arxiv.org/abs/2409.10164
Code Repository: https://github.com/Nicolinho/QRM
Method Overview: QRM generates a distribution over rewards by aggregating individual distributions over attribute scores like helpfulness and harmlessness.

This model uses Skywork/Skywork-Reward-Llama-3.1-8B-v0.2 as backbone and used Skywork/Skywork-Reward-Preference-80K-v0.2 for training the gating network. Apart from this, it has been trained exactly as described in the tech report.

Demo Code

# export ACCELERATE_MIXED_PRECISION=bf16
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer
device = "cuda"
path = "nicolinho/QRM-Llama3.1-8B-v2"
model = AutoModelForSequenceClassification.from_pretrained(path, torch_dtype=torch.bfloat16, device_map=device, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(path, use_fast=True)
# We load a random sample from the validation set of the HelpSteer dataset
prompt = 'Does pineapple belong on a Pizza?'
response = "There are different opinions on this. Some people like pineapple on a Pizza while others condemn this."
messages = [{"role": "user", "content": prompt},
           {"role": "assistant", "content": response}]
input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt").to(device)
with torch.no_grad():
   output = model(input_ids)
   # Expectation of the reward distribution
   reward = output.score.cpu().float() 
   # Quantile estimates for the quantiles 0.05, 0.1, ..., 0.9, 0.95 representing the distribution over rewards
   reward_quantiles = output.reward_quantiles.cpu().float()

# The attributes of the 5 reward objectives
attributes = ['helpsteer-helpfulness','helpsteer-correctness','helpsteer-coherence',
   'helpsteer-complexity','helpsteer-verbosity']

Citation

If you find this work useful for your research, please consider citing:

@article{dorka2024quantile,
  title={Quantile Regression for Distributional Reward Models in RLHF},
  author={Dorka, Nicolai},
  journal={arXiv preprint arXiv:2409.10164},
  year={2024}
}