MOOSE-Star-HC-R1D-7B: Scientific Hypothesis Generation

MOOSE-Star-HC-R1D-7B is a 7 billion parameter model, based on DeepSeek-R1-Distill-Qwen-7B, specialized in generating scientific hypotheses. Developed by Zonglin Yang and Lidong Bing, this model is fine-tuned for incremental hypothesis composition, a critical task in scientific discovery. It operates by taking a research question, background survey, and a new inspiration paper (title + abstract) to output a "delta hypothesis."

Key Capabilities

• Structured Hypothesis Generation: Deconstructs hypotheses into four components: Inspiration, Motivation (WHY), Mechanism (HOW IT WORKS), and Methodology (HOW IT'S INTEGRATED).
• Incremental Composition: Designed to integrate new research insights with existing knowledge to refine or expand hypotheses.
• Scientific Discovery Workflow: Optimized for use in workflows where hypotheses are built up iteratively.
• Performance: Outperforms its base model, DeepSeek-R1-Distill-Qwen-7B, in rubric-based evaluations for hypothesis composition, achieving a total score of 4.74 compared to the base model's 4.05.

Training Details

The model was trained using full-parameter SFT (ZeRO-3) on 114,548 samples from the TOMATO-Star-SFT-Data-R1D-32B HC split, with a cutoff length of 8192 tokens. It utilizes a deepseekr1 chat template.

Usage

Users can integrate this model by cloning the MOOSE-Star repository to access necessary prompt templates and utilities. The model is designed to be used with a specific prompt format that guides the generation of structured delta hypotheses.