Data Sources
Every dataset is indexed from its official release channel — hosting mirrors, papers and changelogs. Metadata is refreshed automatically and verified by hand before it enters the leaderboard.
Benchmark index
Dataset Leaderboard
Open training datasets ranked by normalized benchmark performance across quality, coverage and recency.
Summary Metrics
Datasets
1,248
Benchmarks
28
Avg Score
72.4
Last Updated
May 12, 2024
Filters & Sort
Sort by
Ranking Table
| Select for comparison | Rank | Dataset | Overall Score | Quality | Coverage | Recency | Actions |
|---|
About the index
Methodology
How DataRank collects benchmark results, normalizes scores and ranks the 1,248 datasets in the index.
Benchmark Definitions
The index tracks 28 public benchmarks spanning vision, language and multimodal tasks. Each benchmark pins a fixed evaluation split and metric so scores stay comparable across releases.
Scoring
Raw benchmark results are normalized to a 0–100 scale against the field, then averaged into Quality, Coverage and Recency dimensions. The overall score is the weighted mean of the three.
Ranking Rules
Datasets are ranked by overall score, with ties broken by quality. A dataset must report at least five benchmarks in the trailing year to be eligible for ranking.
Update Cadence
Leaderboard scores recompute nightly and benchmark definitions are reviewed quarterly. Major dataset releases trigger an out-of-cycle re-evaluation within 72 hours.
Limitations
Scores summarize public benchmark results and can lag private or unpublished evaluations. Treat rankings as a starting point for due diligence, not a substitute for it.