ZonglinY/MOOSE-Star-R1D-7B

MOOSE-Star-R1D-7B: Multi-Task Scientific Discovery Model

MOOSE-Star-R1D-7B (MS-7B) is a 7 billion parameter language model developed by ZonglinY, specifically fine-tuned for scientific discovery. Built upon the DeepSeek-R1-Distill-Qwen-7B base model, MS-7B is unique in its ability to perform two distinct, yet complementary, tasks within scientific research: inspiration retrieval (IR) and hypothesis composition (HC).

Key Capabilities

• Inspiration Retrieval: Identifies the most relevant cross-paper inspiration from 15 candidates, including hard negatives, to solve fundamental problems in scientific research. It achieves 54.34% accuracy, matching dedicated single-task IR models.
• Hypothesis Composition: Generates structured delta hypotheses from inspiration papers, detailing motivation, mechanism, and methodology. This multi-task model significantly outperforms all single-task HC variants, demonstrating robust performance even under varying levels of inspiration noise.
• Unified Multi-task Performance: Successfully integrates both IR and HC capabilities into a single model without compromising IR performance, while substantially improving HC quality.

Good for

• Accelerating Scientific Discovery: Designed to assist researchers in identifying novel connections and formulating new hypotheses.
• Complex Scientific Reasoning: Its multi-task nature makes it suitable for workflows requiring both information synthesis and creative hypothesis generation.
• Robustness: Shows improved performance in hypothesis composition, particularly in challenging scenarios with noisy inspiration data, suggesting enhanced reasoning transfer between tasks.