“`html
A British researcher has shared a significant collection on Hugging Face, featuring an 7B model post-trained for scientific hypothesis discovery and the dataset behind it. The paper was accepted at ICML 2026.
- The collection includes MS-IR-7B, MS-HC-7B, and MS-7B models tailored for inspiration retrieval, hypothesis composition, and joint use, respectively. These are built on top of the base model DeepSeek-R1-Distill-Qwen-7B.
- TOMATO-Star is a 108,717 paper dataset decomposed into (background, hypothesis, inspirations), with each inspiration anchored to a real citation. The dataset covers various scientific fields such as biology, chemistry, medicine, medical imaging, psychology, and cognitive science.
- The model has been evaluated on its ability to retrieve scientific hypotheses. MS-7B and MS-IR-7B both outperformed previous models in terms of inspiration retrieval accuracy, with MS-7B achieving a 54.34% success rate compared to the base model’s 28.42%.
This work is significant as it provides a robust framework for AI systems to understand and generate scientific hypotheses, potentially aiding in fields such as drug discovery, disease diagnosis, and more.
“`
– Takeaways:
– The collection offers a new approach to using large language models (LLMs) for scientific hypothesis generation.
– It includes detailed metrics on model performance, which can be used to benchmark other similar systems.
– This work could have significant applications in various scientific research areas by automating or enhancing parts of the discovery process.
