README.md

metadata

base_model:
  - CodeGoat24/UnifiedReward-qwen-7b
datasets:
  - CodeGoat24/HPD
  - CodeGoat24/OIP
  - CodeGoat24/EvalMuse
  - CodeGoat24/ShareGPTVideo-DPO
  - CodeGoat24/LLaVA-Critic-113k
  - CodeGoat24/VideoDPO
  - CodeGoat24/Text-2-Video-Human-Preferences
  - CodeGoat24/OpenAI-4o_t2i_human_preference
  - CodeGoat24/ImageGen_Reward_Cold_Start
license: mit
library_name: transformers
pipeline_tag: image-text-to-text

Model Summary

Unified-Reward-Think-qwen-7b is a unified multimodal CoT reward model, capable of multi-dimensional, step-by-step long-chain reasoning for both visual understanding and generation reward tasks. This model serves as the pairwise preference reward model for the framework presented in the paper Pref-GRPO: Pairwise Preference Reward-based GRPO for Stable Text-to-Image Reinforcement Learning.

For further details on Pref-GRPO and this reward model, please refer to the following resources:

• 📰 Paper: Pref-GRPO: Pairwise Preference Reward-based GRPO for Stable Text-to-Image Reinforcement Learning
• 🪐 Project Page: https://codegoat24.github.io/UnifiedReward/Pref-GRPO
• 💻 GitHub Repository (Pref-GRPO framework): https://github.com/CodeGoat24/Pref-GRPO
• 🤗 Model Collections: https://huggingface.co/collections/CodeGoat24/unifiedreward-models-67c3008148c3a380d15ac63a
• 🤗 Dataset Collections: https://huggingface.co/collections/CodeGoat24/unifiedreward-training-data-67c300d4fd5eff00fa7f1ede
• 👋 Point of Contact: Yibin Wang

Quick Start

All inference codes for using this reward model are provided in our github sub-directory.

We take image understanding assessment as example here:

import json
import random
import torch
import tqdm
from PIL import Image
import warnings
import os
import requests # Added for fetching image from URL
from transformers import AutoProcessor, AutoTokenizer, Qwen2_5_VLForConditionalGeneration
from qwen_vl_utils import process_vision_info

warnings.filterwarnings("ignore")

model_path = "CodeGoat24/UnifiedReward-Think-qwen-7b"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_path, torch_dtype="auto", device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_path)


url = "https://github.com/LLaVA-VL/blog/blob/main/2024-10-03-llava-critic/static/images/critic_img_seven.png?raw=True"
image = Image.open(requests.get(url, stream=True).raw)

Query = 'What does this image present?'
R1 = 'The image is a black and white sketch of a line that appears to be in the shape of a cross. The line is a simple and straightforward representation of the cross shape, with two straight lines intersecting at a point.'
R2 = 'This is a handwritten number seven.'

prompt_text = ("Given a question and a reference image, please analyze in detail the two provided answers (Answer 1 and Answer 2). " \
            "Evaluate them based on the following three core dimensions:
" \
            "1. Semantic accuracy: How well the answer reflects the visual content of the image
" \
            "2. Correctness: Whether the answer is logically and factually correct
" \
            "3. Clarity: Whether the answer is clearly and fluently expressed
" \
            "You may also consider additional dimensions if you find them relevant (e.g., reasoning ability, attention to detail, multimodal grounding, etc.). " \
            "For each dimension, provide a score from 1 to 10 for both answers, and briefly explain your reasoning. " \
            "Then, compute the total score for each answer by explicitly adding the scores for all dimensions and showing the full calculation. " \
            "Enclose your full reasoning within <think> and </think> tags. " \
            "Then, in the <answer> tag, output exactly one of the following: 'Answer 1 is better' or 'Answer 2 is better'. No other text is allowed in the <answer> section.

" \
            "Example format:
" \
            "<think>
" \
            "1. Semantic accuracy: Answer 1 (9/10) - ...; Answer 2 (7/10) - ...
" \
            "2. Correctness: Answer 1 (8/10) - ...; Answer 2 (7/10) - ...
" \
            "3. Clarity: Answer 1 (9/10) - ...; Answer 2 (8/10) - ...
" \
            "[Additional dimensions if any]: Answer 1 (6/10) - ...; Answer 2 (7/10) - ...
" \
            "Total score:
Answer 1: 9+8+9+6=32
Answer 2: 7+7+8+7=29
" \
            "</think>
" \
            "<answer>Answer 1 is better</answer>

" \
            "**Note: In the example above, scores and the final answer are placeholders meant only to demonstrate the format. Your actual evaluation should be based on the quality of two given answers.**

"
            f"Your task is provided as follows:
Question: [{Query}]
Answer 1: [{R1}]
Answer 2: [{R2}]")

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": prompt_text},
        ],
    }
]

chat_input = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)

inputs = processor(
    text=[chat_input],
    images=image_inputs,
    videos=video_inputs,
    return_tensors="pt",
    padding=True
).to("cuda")

with torch.no_grad():
    generated_ids = model.generate(**inputs, max_new_tokens=4096)
generated_trimmed = [
    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output = processor.batch_decode(generated_trimmed, skip_special_tokens=True)[0]

print(output)

Citation

@article{Pref-GRPO&UniGenBench,
  title={Pref-GRPO: Pairwise Preference Reward-based GRPO for Stable Text-to-Image Reinforcement Learning},
  author={Wang, Yibin and Li, Zhimin and Zang, Yuhang and Zhou, Yujie and Bu, Jiazi and Wang, Chunyu and Lu, Qinglin, and Jin, Cheng and Wang, Jiaqi},
  journal={arXiv preprint arXiv:2508.20751},
  year={2025}
}