---
language: [en, fr, nl, de, pl, el, ja]
license: cc-by-nc-4.0
pretty_name: TabFix multilingual table error pairs
task_categories: [token-classification]
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.parquet
  - split: validation
    path: data/validation.parquet
  - split: test
    path: data/test.parquet
tags: [tabular-data, error-detection, data-cleaning, synthetic, xml]
---

# TabFix multilingual table error pairs — version 2.0

This release keeps 18 business error categories and separates executable deterministic detection from two residual neural categories: `text.encoding` and `text.spelling`. The same repository and family-disjoint splits are retained.

| Split | Records | Open-vocabulary views |
|---|---:|---:|
| train | 27948 | 3260 |
| validation | 17127 | 1844 |
| test | 32776 | 3540 |

The seven string columns remain `id`, `split`, `family_id`, `clean_xml`, `corrupt_xml`, `errors`, `metadata`. Named split files are in `data/`; `dataset.parquet` combines them.

Metadata adds `neural_error_indices`, `error_routes`, `neural_view` and `correction_candidates` (complete-cell targets, including valid copies). Detection reads only corrupt XML; correction renders `<original>observed cell</original><replacement>[MASK]…</replacement>` inside the selected cell. Targets are never included in inference input. Empty responses use END_EDIT; EDIT_PAD is distinct from batch padding.

Original closed-vocabulary examples remain. Additional spelling/encoding views remove enum/lexicon constraints, retain the authored reference and share their source family/split. These are schema-ablation training augmentations, not newly collected real text. Matching valid copies include unusual names and scripts. No new claim of open-world semantic accuracy is made. Missing values without recoverable contents and linked field swaps are excluded from independent-cell correction supervision.

Regex contracts were checked for accidental double escaping; no normalization was needed. Data values and partitions were preserved. All original sources remain; the dataset is predominantly synthetic. Family holdout is retained; no stronger structural holdout claim is made for the additional schema views.

`audit.json` records current counts, `label_map.json` defines routing and labels, `release.json` gives checksums, and `provenance.json` retains source attribution. Version 1 remains accessible through Git revision history.

## Licensing and attribution

Project-authored synthetic content, annotations and documentation are licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). Commercial use of that content requires permission from the project owner.

USDA FoodData Central data is public domain and published under CC0 1.0, as described in the [FoodData Central API guide](https://fdc.nal.usda.gov/api-guide/). Its source terms remain applicable independently of the project-authored portions. Attribution: **U.S. Department of Agriculture, Agricultural Research Service. FoodData Central, 2019. fdc.nal.usda.gov.** Source-specific provenance and terms are recorded in `provenance.json`.