OpenMOSS-Team/SciJudge-4B-2605

SciJudge-4B-2605: AI for Scientific Paper Evaluation

SciJudge-4B-2605 is a 4 billion parameter model developed by OpenMOSS-Team, built upon the Qwen3-4B-Instruct-2507 architecture. Its core function is to evaluate scientific papers, specifically predicting which of two given papers (with titles, abstracts, and publication dates) will achieve a higher citation count.

Key Capabilities & Features

• Citation Impact Prediction: Specialized in comparing two scientific papers and determining which is likely to have greater future citation impact.
• Contextual Analysis: Utilizes paper titles, abstracts, and publication dates for its comparative judgment.
• Base Model: Fine-tuned from Qwen3-4B-Instruct-2507, indicating a strong foundation in instruction following and general language understanding.
• Training Methodology: Trained using GRPO with DAPO loss and an external preference reward, leveraging 720,341 preference pairs from the SciJudgeBench dataset.
• Performance: Achieves an average accuracy of 77.3% on the SciJudgeBench test split (MAIN_1000 in-domain evaluation set), significantly outperforming its base model (58.1%).

Good For

• Scientific Research Analysis: Ideal for tasks requiring an AI to assess the potential impact or 'taste' of scientific publications.
• Academic Trend Prediction: Useful for researchers or institutions interested in forecasting the influence of new scientific work.
• Benchmarking: Serves as a smaller, efficient model for evaluating scientific judgment tasks, complementing its larger counterpart, SciJudge-30B-2605.