ZonglinY/MOOSE-Star-IR-R1D-7B
ZonglinY/MOOSE-Star-IR-R1D-7B is a 7 billion parameter model, fine-tuned from DeepSeek-R1-Distill-Qwen-7B, specifically designed for scientific hypothesis generation within the MOOSE-Star framework. This model excels at selecting the most relevant cross-paper inspiration from 15 candidates given a research background, outputting chain-of-thought reasoning. It is optimized for identifying solutions to fundamental problems in existing scientific methods, achieving 54.37% accuracy on this task.
Loading preview...
MOOSE-Star-IR-R1D-7B: Scientific Inspiration Retrieval
MOOSE-Star-IR-R1D-7B is a 7 billion parameter model developed by ZonglinY, fine-tuned from the DeepSeek-R1-Distill-Qwen-7B base model. Its core function is to assist in scientific hypothesis generation by identifying the most suitable cross-paper inspiration from a set of 15 candidates, given a specific research background. This model is a key component of the MOOSE-Star framework, which aims to unlock tractable training for scientific discovery.
Key Capabilities
- Inspiration Selection: Selects the single most relevant cross-paper inspiration from 15 diverse candidates (1 ground truth, 14 hard negatives).
- Chain-of-Thought Reasoning: Generates a detailed reasoning process before outputting the selected inspiration and its justification.
- Scientific Problem Solving: Designed to identify papers that can best solve fundamental problems or limitations in current research approaches.
- Hierarchical Search: Intended for use in a hierarchical search pipeline with O(log N) complexity.
Performance
- Achieves 54.37% accuracy in inspiration selection, significantly outperforming its base model (28.42%) and random selection (6.70%).
Good For
- Researchers and developers working on scientific discovery automation.
- Applications requiring automated hypothesis generation or literature-based inspiration retrieval.
- Integrating into systems that need to identify novel connections between scientific papers to address research gaps.