LxzGordon/URM-LLaMa-3.1-8B

• Paper: https://arxiv.org/pdf/2410.00847

• Model: URM-LLaMa-3.1-8B

  • Fine-tuned from Skywork-Reward-Llama-3.1-8B

Architecture

URM is one of the RMs in the figure.

Alignment Results

Results of using uncertainty estimates to improve LLM alignment. Rewards with low uncertainty are more reliable and result in better alignment.

Brief

URM-LLaMa-3.1-8B is an uncertain-aware reward model. This RM consists of a base model and an uncertainty-aware and attribute-specific value head. The base model of this RM is from Skywork-Reward-Llama-3.1-8B.

URM involves two-stage training: 1. attributes regression and 2. gating layer learning.

Attribute Regression

Dataset: HelpSteer2

During training, instead of multi-attributes scores, outputs of the uncertainty-aware value head are parameters of a normal distribution, from which scores are sampled. Then we run regression on the outputs with the labels to train the value head. To enable gradient back-propagation, reparameterization technique is used.

Gating Layer Learning

Dataset: Skywork-Reward-Preference-80K-v0.1

Inspired by ArmoRM, we learn a gating layer to combine the multi-attribute scores instead of the fixed weights in SteerLM-RM. Learning objective of the gating layer is to prioritize chosen responses over rejected responses through the BT loss. We only use the five attributes from HelpSteer2: Helpfulness, Correctness, Coherence, Complexity and Verbosity. During this process, the value head and base model are kept frozen.

Usage

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_name = "LxzGordon/URM-LLaMa-3.1-8B"
model = AutoModelForSequenceClassification.from_pretrained(
    model_name,
    device_map='auto',
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "What is the range of the numeric output of a sigmoid node in a neural network?"
response1 = "The output of a sigmoid node is bounded between -1 and 1."
response2 = "The output of a sigmoid node is bounded between 0 and 1."

resp1 = [{"role": "user", "content": prompt}, {"role": "assistant", "content": response1}]
resp2 = [{"role": "user", "content": prompt}, {"role": "assistant", "content": response2}]

# Format and tokenize the conversations
resp1 = tokenizer.apply_chat_template(resp1, tokenize=False)
resp2 = tokenizer.apply_chat_template(resp2, tokenize=False)
resp1 = tokenizer(resp1, return_tensors="pt").to(model.device)
resp2 = tokenizer(resp2, return_tensors="pt").to(model.device)

with torch.no_grad():
    score1 = model(resp1['input_ids'],attention_mask=resp1['attention_mask']).logits[0][0].item()
    score2 = model(resp2['input_ids'],attention_mask=resp2['attention_mask']).logits[0][0].item()
print(score1,score2)

# Response 1 score: 2.3285412788391113, Response 2 score: 12.438033103942871

Reference

Please cite

@article{lou2024uncertainty,
  title={Uncertainty-aware Reward Model: Teaching Reward Models to Know What is Unknown},
  author={Lou, Xingzhou and Yan, Dong and Shen, Wei and Yan, Yuzi and Xie, Jian and Zhang, Junge},
  journal={arXiv preprint arXiv:2410.00847},
  year={2024}
}