CompassJudger-2

src="https://img.shields.io/badge/CompassJudger--2-Paper-red?logo=arxiv&logoColor=red" alt="CompassJudger-2" style="display: inline-block; vertical-align: middle;" />

Introduction

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.

Key contributions of our work include:

• Advanced Data Strategy: We employ a task-driven, multi-domain data curation and synthesis strategy to enhance the model's robustness and domain adaptability.
• 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.
• 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.
• 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.

This repository contains the CompassJudger-2 series of models, fine-tuned on the Qwen2.5-Instruct series.

Models

Quickstart

Here is a simple example demonstrating how to load the model and use it for pairwise evaluation.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = "opencompass/CompassJudger-2-7B-Instruct"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Example: Pair-wise Comparison
prompt = """
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.

- Do not let the order of presentation, response length, or assistant names influence your judgment.
- Base your decision solely on how well each response addresses the user’s question and adheres to the instructions.

Your final reply must be structured in the following format:
{
  "Choice": "[Model A or Model B]"
}

User Question: {question}

Model A's Response: {answerA}

Model B's Response: {answerB}

Now it's your turn. Please provide selection result as required:
"""

messages = [
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=2048
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

Evaluation

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.

For detailed benchmark performance and methodology, please refer to our 📑 [Paper](https://arxiv.org/a