---
license: other
pretty_name: DSEvaluation Data
---

# DSEvaluation Data

Versioned data mirror and access catalog for all **9** benchmarks registered in
[DSEvaluation](https://github.com/yangzhr1/DSEvaluation).

| Directory | Benchmark | Tasks | Public content in this repository | Raw data |
| --- | --- | ---: | --- | --- |
| `agenticdatabench/` | AgenticDataBench | 246 | Tasks, datasets, gold data, evaluator | Mirrored |
| `codabench/` | CoDA-Bench | 1,009 | Task metadata, license, archive checksums | Official Kaggle-derived archives |
| `dabench/` | DABench | 257 | Questions, tables, labels | Mirrored |
| `dabstep/` | DABstep | 450 | Context, tasks, license | Mirrored; historical submissions excluded |
| `dacode/` | DA-Code | 500 | Source, gold data, evaluation configs | Mirrored |
| `longds/` | LongDS | 68 | Pinned source manifest | Official dataset |
| `mledojo/` | MLE-Dojo | 10 | Selected task catalog and access instructions | Official Kaggle competitions |
| `moscibench/` | MoSciBench | 88 | Official archives, manifest, license | Mirrored with upstream provenance |
| `sab/` | ScienceAgentBench | 102 | Verified task metadata and access instructions | Official password-protected archive |

The top-level directory set intentionally matches the 9 benchmark plugins.
`metadata-mirrored` entries are runnable through DSEvaluation, but their raw
assets remain on official providers because upstream redistribution terms do
not permit a public copy here. In particular, ScienceAgentBench asks users not
to redistribute unzipped data, and MLE-Dojo requires accepting each Kaggle
competition's terms.

Prepare any benchmark through the same CLI:

```bash
uv run dseval prepare BENCHMARK
```

See `catalog.json` for machine-readable availability and source information,
and `manifest.json` for file sizes and SHA-256 checksums.