OpenMOSS-Team/SciJudge-4B
SciJudge-4B is a 4 billion parameter Qwen3-Instruct model developed by OpenMOSS-Team, specifically fine-tuned for scientific paper evaluation. It predicts the higher citation impact between two papers based on their titles, abstracts, and publication dates. This model utilizes a GRPO training method with DAPO loss and an external preference reward for citation-based pairwise judgment, making it specialized for assessing scientific 'taste' and potential influence.
Loading preview...
SciJudge-4B: Scientific Paper Evaluation Model
SciJudge-4B, developed by OpenMOSS-Team, is a specialized 4 billion parameter language model built upon the Qwen3-4B-Instruct-2507 architecture. Its primary function is to evaluate scientific papers and predict which of two given papers will achieve higher citation impact. This is achieved by analyzing their titles, abstracts, and publication dates.
Key Capabilities
- Citation Impact Prediction: Determines which of two research papers is likely to receive more citations.
- Scientific 'Taste' Assessment: Trained to understand and judge the potential influence and relevance of scientific work.
- Contextual Analysis: Utilizes paper titles, abstracts, and publication dates for its evaluation.
Training and Methodology
The model was fine-tuned using a GRPO (Generalized Reinforcement Learning from Human Feedback with Preference Optimization) method, incorporating DAPO (Direct Preference Optimization with Auxiliary Losses). It leverages an external preference reward system specifically designed for citation-based pairwise judgments, distinguishing it from general-purpose LLMs. The training was conducted in bfloat16 precision with a KL coefficient of 0.03.
Use Cases
SciJudge-4B is ideal for researchers, institutions, or platforms that need to:
- Assess the potential impact of scientific publications.
- Aid in curating or ranking research papers.
- Explore AI's capability in understanding and predicting scientific trends.
For current experiments, users are recommended to use the newer release, SciJudge-4B-2605.