MOOSE-Star 7B Released

The MOOSE-Star collection has been released on Hugging Face. A 7B model post-trained for scientific hypothesis discovery is provided along with its underlying dataset, and the paper has been accepted to ICML 2026.

• The models consist of MS-IR-7B (inspiration retrieval), MS-HC-7B (hypothesis composition), and MS-7B (joint IR+HC).
• All are based on DeepSeek-R1-Distill-Qwen-7B, and are divided according to reasoning, retrieval, and hypothesis composition tasks.

TOMATO-Star is a dataset that decomposes 108,717 NCBI papers into background / hypothesis / inspirations.

• Each inspiration is linked to an actual paper citation, increasing traceability.
• Application areas include biology, chemistry, medicine, medical imaging, psychology, cognitive science.
• About 38,400 A800 GPU-hours were used for data construction preprocessing.

Evaluation was conducted with a strict temporal split of train ≤ September 2025, test = October 2025.

• Inspiration retrieval accuracy was reported as MS-IR-7B 54.37%, MS-7B 54.34%.
• Comparison values were Random 6.70%, base 28.42%, Claude Sonnet 4.6 45.02%, DeepSeek-R1 45.11%, Gemini-3 Flash 51.44%, GPT-5.4 51.50%, Gemini-3 Pro 54.89%.

On the execution side, it can be used directly in any environment where DeepSeek-R1-Distill-Qwen-7B runs, and llama.cpp / vLLM / SGLang are also supported.

• Memory requirements are about 14GB on fp16.
• It is within the range that can run even on a 24GB single GPU environment.
• The code license is Apache-2.0, and the data is CC-BY-4.0.