SciCloze-900: GCSE Combined Science Benchmark for Language Models

SciCloze-900 is a cloze-style benchmark containing 900 multiple-choice items for evaluating small base language models. The benchmark includes 300 biology questions, with chemistry and physics items making up the remainder. It was created by veyra-ai and last updated on June 14, 2026.

Use Cases

• Benchmarking model accuracy on science questions based on the 900 cloze items
• Comparing model performance across biology, chemistry, and physics subjects based on the described subject breakdown
• Analyzing the relationship between model parameter count and benchmark score based on the provided frontier graph

Strengths

• Contains 900 total multiple-choice cloze items for structured evaluation
• Includes a subject breakdown with 300 biology questions, providing domain-specific analysis
• Provides benchmark frontier data comparing model accuracy with parameter count

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• The benchmark only lists specific evaluated models; higher-scoring models may exist

Provenance

Source

veyra-ai

Freshness

Last updated 2026-06-14 03:44:50