---
license: other
license_name: mixed-sources-see-card
pretty_name: ARB ML and data science raw corpus
language:
- en
tags:
- data-science
- machine-learning
- statistics
- causal-inference
- bioinformatics
- tabular
- documentation
size_categories:
- 10K<n<100K
---

# ARB ML and data science raw corpus

Study and practice material for the **ML and data science** track of Applied RSI Bench
(ARB): the DARE-bench, QRData and DSBio tasks. An ARB post-training agent gets this
corpus read-only before its model is evaluated. It holds datasets to practice on, the
three tasks' container images (to run analyses offline), and library and methodology
documentation.

The corpus holds **no held-out test case**. Every dataset was checked against a private
index of the three benchmarks' test cases (511 questions, 1,092 test tables): a dataset is
refused if it carries a test dataset's identity or file, a whole test table (in any row
and column order), 20 % of a test table's distinct rows, or 2 of its distinctive columns.
Every text file, image layers included, was refused if it shares a 13-word run (at least
6 of the words with a letter) with a test question, or links to the sites of the three
textbooks that 86 of QRData's 155 test questions come from (OpenIntro Statistics,
Quantitative Social Science, Causal Inference for the Brave and True). DARE-bench folders keep only `metadata.txt` and the training table, never the
validation files, labels or questions.

| file | files | unpacked → archive | tree_sha256 |
|---|---|---|---|
| `L.tar.gz` | 17,542 | 2.62 GB → 2.33 GB | `2d9bc63a53bd823eb6daf8605db3a3ecbc1c4cf2720f61a1e0996feddbc1344f` |

`L.tar.gz` sha256: `484aafa5f7075cab0f4c9b8aaceaa32a81d11788eb9a74da285b3ced0ffa6ec9`

## What is inside

| folder | what | size |
|---|---|---|
| `docs/` | documentation pages (with docs/PACKET.json and docs/README.md) (17,467 files) | 323.3 MB |
| `datasets/` | 16 datasets: original files, dataset.json, verification.json (64 files) | 74.8 MB |
| `images/` | OCI image archives, catalog and recipes (9 files) | 2.22 GB |

- `datasets/<id>/data/`: each dataset's original files, unchanged. `dataset.json` gives the
  source, license, identities and each file's format and schema; `verification.json` which
  files loaded, and in which image.
- `images/`: the three tasks' environments as OCI archives (`podman load -i
  images/<id>.oci.tar`): `dare-task` (DARE-bench), `dsgym-instance` (QRData, from DSGym)
  and `dsbio-task` (DSBio), built for linux/amd64. `catalog.json` lists them. Each one read
  the datasets with no network: `dsbio-task` all 32 files, the other two 30 (no h5py for
  the two `.h5ad` files).
- `docs/`: 17,465 documentation pages as text, one folder per source host;
  `docs/PACKET.json` gives each page's source URL. Python 3.11 and 3.12 and pandas 2.2.3
  come from their documentation archives.

`PACKET.json` lists every file with its sha256 and size; `tree_sha256` is the sha256 of
the sorted lines `"{path} {sha256}\n"`. `RELEASE.json` and `catalog.json` record the
build inputs. `RELEASE.json` says `ready: false` with a list of pending checks: the
build tool always writes that, because it cannot certify a release by itself.

## Use

```bash
hf download junlinw/arb-ml-data-science-data-packet L.tar.gz --repo-type dataset --revision <commit> --local-dir .
mkdir -p ml-data-science && tar -xzf L.tar.gz -C ml-data-science
```

ARB pins the exact commit it installs in the tasks' `benchmark.toml`.

## Datasets and licenses

| id | title | license | source |
|---|---|---|---|
| dare-varshitanalluri-crop-price-prediction-dataset-reg | Crop Price Prediction Dataset | MIT | [Kaggle varshitanalluri/crop-price-prediction-dataset](https://www.kaggle.com/datasets/varshitanalluri/crop-price-prediction-dataset) via DARE-bench |
| dare-harshvir04-index-value-economic-dataset-reg | Index value (economic dataset) | MIT | [Kaggle harshvir04/index-value-economic-dataset](https://www.kaggle.com/datasets/harshvir04/index-value-economic-dataset) via DARE-bench |
| dare-vivianyfwang-capital-bike-share-2018-20218-reg | Capital Bike Share | CC0-1.0 | [Kaggle vivianyfwang/capital-bike-share-2018-20218](https://www.kaggle.com/datasets/vivianyfwang/capital-bike-share-2018-20218) via DARE-bench |
| dare-ahsan81-used-handheld-device-data-reg | Used Phones & Tablets Pricing Dataset | CC0-1.0 | [Kaggle ahsan81/used-handheld-device-data](https://www.kaggle.com/datasets/ahsan81/used-handheld-device-data) via DARE-bench |
| dare-himanshu9648-crop-price-prediction-dataset-in-india-ts | Crop Price Prediction in India dataset | CC0-1.0 | [Kaggle himanshu9648/crop-price-prediction-dataset-in-india](https://www.kaggle.com/datasets/himanshu9648/crop-price-prediction-dataset-in-india) via DARE-bench |
| dare-olegshpagin-visa-stock-price-prediction-dataset-ts | Visa Stock Price Prediction Dataset | CC0-1.0 | [Kaggle olegshpagin/visa-stock-price-prediction-dataset](https://www.kaggle.com/datasets/olegshpagin/visa-stock-price-prediction-dataset) via DARE-bench |
| dare-negmgh-google-2020-2025-stock-market-ts | Google 2020-2025 Stock Market | Apache-2.0 | [Kaggle negmgh/google-2020-2025-stock-market](https://www.kaggle.com/datasets/negmgh/google-2020-2025-stock-market) via DARE-bench |
| dare-pranav941--world-food-wealth-bank-ts | 🍕Food Bank🏦of the World🌍 | CC0-1.0 | [Kaggle pranav941/-world-food-wealth-bank](https://www.kaggle.com/datasets/pranav941/-world-food-wealth-bank) via DARE-bench |
| dare-miadul-brain-tumor-dataset-class | Brain Tumor Dataset | MIT | [Kaggle miadul/brain-tumor-dataset](https://www.kaggle.com/datasets/miadul/brain-tumor-dataset) via DARE-bench |
| dare-ashishgup-netflix-rotten-tomatoes-metacritic-imdb-class | Latest Netflix data with 26+ joined attributes | CC0-1.0 | [Kaggle ashishgup/netflix-rotten-tomatoes-metacritic-imdb](https://www.kaggle.com/datasets/ashishgup/netflix-rotten-tomatoes-metacritic-imdb) via DARE-bench |
| dare-mexwell-image-sentiment-polarity-class | 🖼️ Image Sentiment Polarity | CC0-1.0 | [Kaggle mexwell/image-sentiment-polarity](https://www.kaggle.com/datasets/mexwell/image-sentiment-polarity) via DARE-bench |
| dare-thedevastator-hunt-prices-for-north-american-mammals-class | Hunt Prices for North American Mammals | CC0-1.0 | [Kaggle thedevastator/hunt-prices-for-north-american-mammals](https://www.kaggle.com/datasets/thedevastator/hunt-prices-for-north-american-mammals) via DARE-bench |
| uci-wine-quality | Wine Quality | CC-BY-4.0 | [archive.ics.uci.edu](https://archive.ics.uci.edu/dataset/186/wine+quality) |
| uci-adult | Adult (census income) | CC-BY-4.0 | [archive.ics.uci.edu](https://archive.ics.uci.edu/dataset/2/adult) |
| cellxgene-4d4c7412 | Mouse Craniofacial Expression Cranial Neural Crest | CC-BY-4.0 | [cellxgene.cziscience.com](https://cellxgene.cziscience.com/collections/690fb28b-e364-4d3d-8934-bae50ae93857) |
| cellxgene-6dfbbfae | E9.0: Puck_220718_11 | CC-BY-4.0 | [cellxgene.cziscience.com](https://cellxgene.cziscience.com/collections/d74b6979-efba-47cd-990a-9d80ccf29055) |

The DARE-bench folders come from the train split of `Snowflake/dare-bench` (commit
`7f202a10`); each one is a Kaggle dataset, and the license is the one its Kaggle page gave
on 2026-10-26.

## Documentation sources and licenses

Only pages whose publisher allows redistribution are included; 81 pages from
other hosts (for example gsea-msigdb.org, otexts.com and mixtape.scunning.com) were left
out. Pages were converted from HTML to text, which is a change to the original. Each
page's source URL is in `docs/PACKET.json`.

| license | pages | sources (pages) |
|---|---|---|
| BSD-3-Clause | 11,184 | pandas.pydata.org (6,357), numpy.org (1,000), networkx.org (1,000), docs.scipy.org (1,000), scikit-learn.org (969), scanpy.readthedocs.io (309), www.statsmodels.org (238), squidpy.readthedocs.io (139), anndata.readthedocs.io (125), umap-learn.readthedocs.io (30), gseapy.readthedocs.io (8), github.com/broadinstitute/tensorqtl (7), docs.h5py.org (2) |
| PSF-2.0 | 2,120 | docs.python.org (2,120) |
| Apache-2.0 | 1,366 | arrow.apache.org (1,000), www.ensembl.org (132), xgboost.readthedocs.io (109), github.com/theislab/single-cell-best-practices (55), github.com/openvax/pyensembl (27), openvax.github.io (22), causalml.readthedocs.io (21) |
| MIT | 957 | www.pywhy.org (397), pgmpy.org (291), lingam.readthedocs.io (65), bedtools.readthedocs.io (43), daler.github.io (35), causal-learn.readthedocs.io (31), pyranges.readthedocs.io (28), github.com/py-why/EconML (25), zarr.readthedocs.io (23), lightgbm.readthedocs.io (14), github.com/swolock/scrublet (3), github.com/chanzuckerberg/single-cell-curation (2) |
| NCSA | 909 | bashtage.github.io (909) |
| us-government-work | 883 | www.itl.nist.gov (883) |
| BSD-2-Clause | 33 | github.com/intro-stat-learning/ISLP_labs (18), patsy.readthedocs.io (15) |
| GPL-3.0 | 13 | leidenalg.readthedocs.io (7), github.com/slowkow/harmonypy (6) |

## Images

The images are built from the tasks' own Dockerfiles: a Python base and the Python and
system packages those files install, each under its own license.

The corpus is shared for research on post-training agents. To have something removed,
open a discussion on this dataset.