---
language:
  - en
license: apache-2.0
base_model: Qwen/Qwen3-4B-Instruct-2507
tags:
  - scientific-evaluation
  - citation-prediction
  - preference-learning
  - GRPO
pipeline_tag: text-generation
library_name: transformers
---

# SciJudge-Qwen3-4B

SciJudge-Qwen3-4B is a fine-tuned language model for **scientific paper evaluation**. Given two academic papers' metadata (title, abstract, publication date), it predicts which paper has a higher citation count — serving as a proxy for assessing research impact and "scientific taste."

This model is part of the paper: **[AI Can Learn Scientific Taste](https://arxiv.org/abs/2603.14473)**.

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "OpenMOSS-Team/SciJudge-4B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="bfloat16", device_map="auto")

messages = [
    {"role": "system", "content": "You are a helpful assistant. You first think about the reasoning process in your mind and then provide the user with the answer."},
    {"role": "user", "content": "Today is 2025-12-10. Based on the titles, abstracts, and publication dates of the following two papers A and B, determine which paper has a higher citation count.\nShow your reasoning process in <reason> </reason> tags. And return the final answer in <answer> </answer> tags. The final answer should contain only 'A' or 'B'.\n\nPaper A:\nTitle: ...\nAbstract: ...\nDate: ...\n\nPaper B:\nTitle: ...\nAbstract: ...\nDate: ..."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.7, top_p=0.8, top_k=20)
response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
print(response)
```

## Training Details

- **Base model:** Qwen3-4B-Instruct-2507
- **Training method:** GRPO (Generative Reward Policy Optimization) with DAPO loss
- **Training data:** 720,341 preference pairs from arXiv papers
- **Learning rate:** 8e-7 (cosine schedule, 5% warmup)
- **Batch size:** 8 per device × 64 GPUs × 2 gradient accumulation = 1024 effective
- **Optimizer:** AdamW (β1=0.9, β2=0.95, weight decay=0.1)
- **Precision:** bfloat16
- **KL coefficient (β):** 0.03

## Citation

```bibtex
@article{scijudge2025,
  title={AI Can Learn Scientific Taste},
  year={2025}
}
```