Knowledge Commonsense Mix: A Fixed Evaluation Set for AI Benchmarks

A fixed evaluation-set composition built with BenchPress, last updated July 7, 2026. The dataset is a mix of established AI benchmarks, sourcing data from repositories like MMLU, ARC Challenge, and HellaSwag at load time. It was created by user seonghyeon0408 and focuses on abilities like evidence retrieval and abstract pattern induction.

Use Cases

• Benchmarking language model performance on evidence retrieval tasks based on the described ability.
• Evaluating abstract pattern induction capabilities in AI systems based on the described ability.
• Comparing model performance across multiple established benchmarks like MMLU and ARC Challenge based on the source list.

Strengths

• Composed of established benchmarks like MMLU and ARC Challenge, providing a known reference point.
• Designed for specific abilities like evidence retrieval and abstract pattern induction, offering a targeted evaluation.

Limitations

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

Provenance

Source

Aggregated from multiple open benchmark repositories on Hugging Face (e.g., tinyBenchmarks/tinyMMLU, allenai/ai2_arc).

Collection Method

Built with BenchPress; source data is streamed at load time from original repositories.

Freshness

Last updated 2026-07-07 18:10:01; freshness should be verified.