SGI-Bench: Scientist-Aligned Benchmark for Evaluating LLMs Across 10 Disciplines

SGI-Bench is a benchmark for evaluating Scientific General Intelligence in large language models. It contains more than 1,000 expert-curated samples spanning 10 scientific disciplines, aligned with the full inquiry cycle. The dataset was created by InternScience and last updated on 2026-06-02.

Use Cases

• Benchmarking LLM performance on scientific inquiry workflows based on the described Deliberation, Conception, Action, and Perception cycle.
• Training or fine-tuning models for scientific reasoning based on the expert-curated samples from 10 disciplines.
• Developing agentic evaluation frameworks for AI based on the described scientist-aligned methodology.

Strengths

• More than 1,000 expert-curated samples provide a substantial evaluation corpus.
• Benchmark spans 10 distinct scientific disciplines for broad coverage.
• Evaluation framework is designed to assess the full agentic inquiry cycle.

Limitations

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

Provenance

Source

InternScience via Hugging Face.

Collection Method

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

Freshness

Last updated 2026-06-02 12:01:45.