Vintage CORE: Two 20-Task Derivatives of the DCLM CORE Evaluation Bundle

Two validated derivatives of the DCLM CORE evaluation bundle, each containing 47,092 rows. The dataset, authored by jbduran, is a benchmark artifact with a last recorded update in July 2026. The original CORE data is not included and must be retrieved from an external source.

Use Cases

• Benchmarking language model performance based on the 20-task structure mentioned in the description
• Comparing model outputs between filtered and restyled data versions described in the repository
• Validating evaluation pipelines using a known derivative of the CORE benchmark

Strengths

• Contains two distinct validated derivatives, each with 47,092 rows
• Specifically designed as a 20-task evaluation bundle for benchmarking
• Last updated on 2026-07-20, indicating recent maintenance

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is known for the derivatives but the original data's scale and specifics are not mirrored
• Description metadata is limited; actual data quality requires manual inspection after download

Provenance

Source

Derived from the DCLM CORE evaluation bundle, with the original hosted on an official nanochat AWS URL.

Collection Method

Created as validated filtered and restyled derivatives of the original benchmark.

Freshness

Last updated 2026-07-20 15:42:23; freshness should be verified