haidang2405/tabrepairbench-replayable-corruption-lineage
TabRepairBench: Replayable Corruption Lineage This is a finite, wholly authored synthetic reference dataset for auditing tabular corruption lineage. Public Lineage v1 contains 72 independent groups and 6,912 clean/corrupt cell pairs across three authored structural generator families, six corruption schedules, two severities, and two audit partitions. It is intentionally not presented as real-world data. It makes no claim of real-data representativeness, causal identification… See the full description on the dataset page: https://huggingface.co/datasets/haidang2405/tabrepairbench-replayable-corruption-lineage.
TabRepairBench: Replayable Corruption Lineage
This is a finite, wholly authored synthetic reference dataset for auditing tabular corruption lineage. Public Lineage v1 contains 72 independent groups and 6,912 clean/corrupt cell pairs across three authored structural generator families, six corruption schedules, two severities, and two audit partitions.
It is intentionally not presented as real-world data. It makes no claim of real-data representativeness, causal identification, method superiority, novelty priority, or state-of-the-art performance. No third-party dataset, private predecessor artifact, scraped row, personal record, or model-generated row is an input.
Load a table
Each logical table is an independent subset. The default paired_cells table is the cell-level audit surface.
from datasets import load_dataset
pairs = load_dataset(
"haidang2405/tabrepairbench-replayable-corruption-lineage",
"paired_cells",
split="data",
)The exact schemas and field descriptions are in `release/DATA_DICTIONARY.json`. CSV copies of the three main tables are preserved for portability, while the viewer uses the typed Parquet copies.
Integrity and semantic replay
The portable verifier reconstructs all published semantics and runs a predeclared mutation suite. After downloading the repository, run:
python release/verify_release.py \
--bundle release \
--mode distributed \
--jsonExpected release ID: tabrepair-public-lineage-v1-98e9cc4162b4a02ef0c3724412d1e5fb5fc04b75c538e416ff8555c04c2d2b3e. The expected replay rate is exactly 1/1 with zero failed checks.
The directory release/ is a byte-identical copy of the public Kaggle bundle. The later publication authorization is external to the immutable generated bundle; this HF mirror was explicitly authorized by the dataset owner on 2026-08-27.
Intended use and limitations
Use this release to test lineage reconstruction, corruption-mechanism routing, schema handling, audit tooling, and reproducible repair evaluation. Do not use its rates as population estimates, train a model and imply real-world utility, or treat its SCMs as empirical causal models. Read `release/DATA_CARD.md`, `release/LIMITATIONS.md`, and `release/README.md` before analysis.
Generated scientific artifacts and dataset-specific documentation are CC BY 4.0. The verifier and complete quick-start notebook are Apache-2.0; see `release/LICENSES.md` for the exact license map.