Update README.md
#1
by
iridescentttt
- opened
README.md
CHANGED
|
@@ -1,122 +1,147 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
model = AutoModelForCausalLM.from_pretrained(
|
| 58 |
-
model_name,
|
| 59 |
-
torch_dtype="auto",
|
| 60 |
-
device_map="auto"
|
| 61 |
-
)
|
| 62 |
-
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
# CompassJudger-2
|
| 7 |
+
|
| 8 |
+
<div align="left" style="line-height: 1;">
|
| 9 |
+
<a href="https://github.com/open-compass/CompassJudger" target="_blank" style="margin: 2px;">
|
| 10 |
+
<img alt="Homepage" src="https://img.shields.io/badge/CompassJudger-GitHub-blue?color=1991ff&logo=github&logoColor=white" style="display: inline-block; vertical-align: middle;"/>
|
| 11 |
+
</a>
|
| 12 |
+
<a href="https://arxiv.org/abs/2507.09104" target="_blank" style="margin: 2px;"">
|
| 13 |
+
<img
|
| 14 |
+
src="https://img.shields.io/badge/CompassJudger--2-Paper-red?logo=arxiv&logoColor=red"
|
| 15 |
+
alt="CompassJudger-2"
|
| 16 |
+
style="display: inline-block; vertical-align: middle;"
|
| 17 |
+
/>
|
| 18 |
+
</a>
|
| 19 |
+
<a href="https://huggingface.co/opencompass" target="_blank" style="margin: 2px;">
|
| 20 |
+
<img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-OpenCompass-536af5?color=536af5&logoColor=white" style="display: inline-block; vertical-align: middle;"/>
|
| 21 |
+
</a>
|
| 22 |
+
<a href="https://github.com/open-compass/CompassJudger/blob/main/LICENSE" style="margin: 2px;">
|
| 23 |
+
<img alt="License" src="https://img.shields.io/badge/License-Apache%202.0-f5de53?color=f5de53&logoColor=white" style="display: inline-block; vertical-align: middle;"/>
|
| 24 |
+
</a>
|
| 25 |
+
</div>
|
| 26 |
+
|
| 27 |
+
## Introduction
|
| 28 |
+
|
| 29 |
+
We introduce **CompassJudger-2**, a novel series of generalist judge models designed to overcome the narrow specialization and limited robustness of existing LLM-as-judge solutions. Current judge models often struggle with comprehensive evaluation, but CompassJudger-2 addresses these limitations with a powerful new training paradigm.
|
| 30 |
+
|
| 31 |
+
Key contributions of our work include:
|
| 32 |
+
|
| 33 |
+
- **Advanced Data Strategy:** We employ a task-driven, multi-domain data curation and synthesis strategy to enhance the model's robustness and domain adaptability.
|
| 34 |
+
- **Verifiable Reward-Guided Training:** We supervise judgment tasks with verifiable rewards, guiding the model's intrinsic reasoning through chain-of-thought (CoT) and rejection sampling. A refined margin policy gradient loss further enhances performance.
|
| 35 |
+
- **Superior Performance:** CompassJudger-2 achieves state-of-the-art results across multiple judge and reward benchmarks. Our 7B model demonstrates competitive accuracy with models that are significantly larger.
|
| 36 |
+
- **JudgerBenchV2:** We introduce a new, comprehensive benchmark with 10,000 questions across 10 scenarios, using a Mixture-of-Judgers (MoJ) consensus for more reliable ground truth.
|
| 37 |
+
|
| 38 |
+
This repository contains the **CompassJudger-2** series of models, fine-tuned on the Qwen2.5-Instruct series.
|
| 39 |
+
|
| 40 |
+
## Models
|
| 41 |
+
|
| 42 |
+
| Model Name | Size | Base Model | Download | Notes |
|
| 43 |
+
| :--------------------------------- | :--: | :------------------- | :----------------------------------------------------------: | :-------------------------------------------- |
|
| 44 |
+
| 👉 **CompassJudger-2-7B-Instruct** | 7B | Qwen2.5-7B-Instruct | 🤗 [Model](https://huggingface.co/opencompass/CompassJudger-2-7B-Instruct) | Fine-tuned for generalist judge capabilities. |
|
| 45 |
+
| 👉 **CompassJudger-2-32B-Instruct** | 32B | Qwen2.5-32B-Instruct | 🤗 [Model](https://huggingface.co/opencompass/CompassJudger-2-32B-Instruct) | A larger, more powerful judge model. |
|
| 46 |
+
|
| 47 |
+
## Quickstart
|
| 48 |
+
|
| 49 |
+
Here is a simple example demonstrating how to load the model and use it for pairwise evaluation.
|
| 50 |
+
|
| 51 |
+
```python
|
| 52 |
+
import torch
|
| 53 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 54 |
+
|
| 55 |
+
model_path = "opencompass/CompassJudger-2-7B-Instruct"
|
| 56 |
+
|
| 57 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 58 |
+
model_name,
|
| 59 |
+
torch_dtype="auto",
|
| 60 |
+
device_map="auto"
|
| 61 |
+
)
|
| 62 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 63 |
+
|
| 64 |
+
# Example: Pair-wise Comparison
|
| 65 |
+
prompt = """
|
| 66 |
+
Please act as an impartial judge to evaluate the responses provided by two AI assistants to the user question below. Your evaluation should focus on the following criteria: helpfulness, relevance, accuracy, depth, creativity, and level of detail.
|
| 67 |
+
|
| 68 |
+
- Do not let the order of presentation, response length, or assistant names influence your judgment.
|
| 69 |
+
- Base your decision solely on how well each response addresses the user’s question and adheres to the instructions.
|
| 70 |
+
|
| 71 |
+
Your final reply must be structured in the following format:
|
| 72 |
+
{
|
| 73 |
+
"Choice": "[Model A or Model B]"
|
| 74 |
+
}
|
| 75 |
+
|
| 76 |
+
User Question: {question}
|
| 77 |
+
|
| 78 |
+
Model A's Response: {answerA}
|
| 79 |
+
|
| 80 |
+
Model B's Response: {answerB}
|
| 81 |
+
|
| 82 |
+
Now it's your turn. Please provide selection result as required:
|
| 83 |
+
"""
|
| 84 |
+
|
| 85 |
+
messages = [
|
| 86 |
+
{"role": "user", "content": prompt}
|
| 87 |
+
]
|
| 88 |
+
|
| 89 |
+
text = tokenizer.apply_chat_template(
|
| 90 |
+
messages,
|
| 91 |
+
tokenize=False,
|
| 92 |
+
add_generation_prompt=True
|
| 93 |
+
)
|
| 94 |
+
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
| 95 |
+
|
| 96 |
+
generated_ids = model.generate(
|
| 97 |
+
**model_inputs,
|
| 98 |
+
max_new_tokens=2048
|
| 99 |
+
)
|
| 100 |
+
generated_ids = [
|
| 101 |
+
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
|
| 102 |
+
]
|
| 103 |
+
|
| 104 |
+
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
|
| 105 |
+
print(response)
|
| 106 |
+
```
|
| 107 |
+
|
| 108 |
+
## Evaluation
|
| 109 |
+
|
| 110 |
+
CompassJudger-2 sets a new state-of-the-art for judge models, outperforming general models, reward models, and other specialized judge models across a wide range of benchmarks.
|
| 111 |
+
|
| 112 |
+
| Model | JudgerBench V2 | JudgeBench | RMB | RewardBench | Average |
|
| 113 |
+
| :--------------------------------- | :------------: | :--------: | :-------: | :---------: | :-------: |
|
| 114 |
+
| **7B Judge Models** | | | | | |
|
| 115 |
+
| CompassJudger-1-7B-Instruct | 57.96 | 46.00 | 38.18 | 80.74 | 55.72 |
|
| 116 |
+
| Con-J-7B-Instruct | 52.35 | 38.06 | 71.50 | 87.10 | 62.25 |
|
| 117 |
+
| RISE-Judge-Qwen2.5-7B | 46.12 | 40.48 | 72.64 | 88.20 | 61.61 |
|
| 118 |
+
| **CompassJudger-2-7B-Instruct** | **60.52** | **63.06** | **73.90** | **90.96** | **72.11** |
|
| 119 |
+
| **32B+ Judge Models** | | | | | |
|
| 120 |
+
| CompassJudger-1-32B-Instruct | 60.33 | 62.29 | 77.63 | 86.17 | 71.61 |
|
| 121 |
+
| Skywork-Critic-Llama-3.1-70B | 52.41 | 50.65 | 65.50 | 93.30 | 65.47 |
|
| 122 |
+
| RISE-Judge-Qwen2.5-32B | 56.42 | 63.87 | 73.70 | 92.70 | 71.67 |
|
| 123 |
+
| **CompassJudger-2-32B-Instruct** | **62.21** | **65.48** | 72.98 | **92.62** | **73.32** |
|
| 124 |
+
| **General Models (for reference)** | | | | | |
|
| 125 |
+
| Qwen2.5-32B-Instruct | 62.97 | 59.84 | 74.99 | 85.61 | 70.85 |
|
| 126 |
+
| DeepSeek-V3-0324 | 64.43 | 59.68 | 78.16 | 85.17 | 71.86 |
|
| 127 |
+
| Qwen3-235B-A22B | 61.40 | 65.97 | 75.59 | 84.68 | 71.91 |
|
| 128 |
+
|
| 129 |
+
|
| 130 |
+
For detailed benchmark performance and methodology, please refer to our [📑 Paper](https://arxiv.org/abs/2507.09104).
|
| 131 |
+
|
| 132 |
+
## License
|
| 133 |
+
|
| 134 |
+
This project is licensed under the Apache 2.0 License. See the [LICENSE](https://github.com/open-compass/CompassJudger/blob/main/LICENSE) file for details.
|
| 135 |
+
|
| 136 |
+
## Citation
|
| 137 |
+
|
| 138 |
+
If you find our work helpful, please consider citing our paper:
|
| 139 |
+
|
| 140 |
+
```bibtex
|
| 141 |
+
@article{zhang2025compassjudger,
|
| 142 |
+
title={CompassJudger-2: Towards Generalist Judge Model via Verifiable Rewards},
|
| 143 |
+
author={Zhang, Taolin and Cao, Maosong and Lam, Alexander and Zhang, Songyang and Chen, Kai},
|
| 144 |
+
journal={arXiv preprint arXiv:2507.09104},
|
| 145 |
+
year={2025}
|
| 146 |
+
}
|
| 147 |
+
```
|