SGI-Bench: Scientist-Aligned LLM Evaluation Across 10 Disciplines

SGI-Bench is a scientist-aligned benchmark for evaluating Scientific General Intelligence in large language models. It spans 10 scientific disciplines and contains more than 1,000 expert-curated samples inspired by Science's 125 Big Questions. The dataset was created by InternScience and last updated on Hugging Face in June 2026.

Use Cases

• Benchmarking LLM deliberation skills based on the structured inquiry cycle framework
• Evaluating model performance on conception tasks across 10 scientific disciplines
• Testing agentic action planning and perception in scientific workflows
• Assessing scientific general intelligence on expert-curated samples from major scientific questions

Strengths

• More than 1,000 expert-curated samples provide substantial evaluation material
• Covers 10 distinct scientific disciplines for broad assessment
• Structured around the four-phase inquiry cycle: Deliberation, Conception, Action, and Perception
• Samples are inspired by Science magazine's 125 Big Questions, grounding them in real scientific challenges

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment for large-scale training
• Description metadata is limited; actual data quality requires manual inspection after download

Provenance

Source

Hugging Face dataset repository by author InternScience

Collection Method

Expert-curated samples inspired by Science's 125 Big Questions

Freshness

Last updated 2026-06-02 12:03:01; freshness should be verified