Antix5/tabular-errors-v1

TabFix multilingual table error pairs

Context-rich XML pairs for detecting and correcting tabular data errors. Each example has an identical schema and row selection in clean_xml and corrupt_xml; errors describe exact repairs within decoded corrupt cells. Clean controls and explicit hard negatives have identical XML and empty error lists.

Files and use

dataset.parquet is the consolidated dataset. audit.json reports actual counts, category coverage, language/domain distributions, compound combinations, family-disjoint splits and integrity checks. provenance.json records source provenance, applicable license terms, and attribution. Build fingerprints and generation details are retained in the release metadata.

All seven Parquet columns are strings:

Column	Meaning
id	Deterministic example content identifier
split	train, validation, or test
family_id	Group kept entirely in one split, including translations
clean_xml	Clean reference context
corrupt_xml	Detector input context
errors	JSON list of {row,column,start,end,replacement,category,mechanism}
metadata	JSON object containing provenance, language, domain, supervision scope, sampled original rows, token counts, generation details, and optional negative evidence

import json
import pyarrow.parquet as pq

rows = pq.read_table("dataset.parquet").to_pylist()
train = [row for row in rows if row["split"] == "train"]
example = train[0]
errors = json.loads(example["errors"])
metadata = json.loads(example["metadata"])

A single Parquet file's split column does not automatically create separate Hugging Face splits. Configure split-specific exports when loading the dataset. Do not feed clean_xml, error labels, or generation metadata to the detector.

Context and schema

XML includes the full schema, explicit nullable states (allowed, forbidden, conditional, unknown), declared formatting and contextual rules, and complete selected rows. Selection combines the first rows, neighboring rows and random rows, deduplicated and sorted by original index. Half-open <gap start="..." end="..."/> elements identify omissions. Optional rows are removed if necessary; individual cells are never truncated. Both XML strings fit the 8,192-token context budget, including special tokens.

XML 1.0 character references preserve CR, LF and tabs; <empty/> means an empty string. Missing data is not automatically an error: a required or conditional schema rule is necessary.

Errors, switches and correction

Categories are stable configuration switches:

• Text: text.encoding, text.invisible, text.spelling.
• Formatting: format.number, format.date, format.boolean, format.case, format.whitespace, format.unit.
• Schema: schema.type, schema.identifier, schema.category.
• Context: consistency.unit, consistency.dependency, consistency.temporal.
• Missingness: missing.required, missing.conditional.
• Structure: structure.shift.

The category is the canonical diagnosis used for supervision; mechanisms name how a defect was generated. The taxonomy does not assert that semantic predicates never overlap. A structural swap can require two edits in one category. Compound examples can contain several categories and independently repairable defects in the same cell.

Coordinates are zero-based Unicode codepoints in the decoded final corrupt cell, with half-open [start,end) ranges. They are not XML byte offsets or tokenizer positions. Original row indices and schema column indices locate the cell. Missing values use zero-width insertion spans. Apply all selected repairs right to left within each cell, using their original coordinates. Exclude disabled categories before applying repairs. Character-to-token alignment and fill-in-middle training serialization are downstream training steps.

Generation and hard negatives

Procedural mutations require a declared lexical or relational contradiction, rather than a rare value alone. Contextual fictional scenarios add implicit arithmetic: equivalent mass quantities, per-item versus whole-pack totals, temporal relationships and valid exceptions. Hard negatives include leading zeros, valid mixed scripts, optional empty fields, accepted date/number/boolean representations and allowed whitespace. Explicit hard-negative evidence is stored separately from ordinary matched clean controls in metadata.negative.

The synthetic corpus covers seven languages and eight source domains. This does not imply equal category support or real-world representativeness. Translated variants share a family and cannot cross splits. Source/template-family splitting is stronger than random row splitting but does not remove shared generator conventions; evaluation on independent real data remains necessary.

Real USDA context is not certified universally error-free prose. Its labels supervise injected defects in the declared eligible columns; unrelated cells and unknown semantic constraints are outside that scope. Respect metadata.label_scope and supervised_columns when constructing detector losses. Empty labels on such a record do not certify every context cell.

Current coverage and remaining work

The first build contains 21,810 pairs: 13,846 corrupted examples and 7,964 clean controls, including 585 explicitly evidenced hard negatives. There are 6,463 multi-category compounds, including 156 same-cell compounds. All 18 categories have positive and explicit hard-negative examples. The maximum observed length is 7,951 tokens.

Coverage is incomplete for a final balanced benchmark: spelling has 56 positive examples, casing 26, and whitespace formatting 36; independent held-out families are still needed for these and other rare contextual categories. Only 299 examples use real USDA tables; 21,511 are synthetic. Deduplication leaves a 63.5% corrupted / 36.5% clean mixture, not a chosen production error prior. Language incidence is nearly balanced (normalized entropy 0.9997), while positive category incidence is less balanced (0.8368). These are descriptive diagnostics, not proof of training sufficiency.

Coverage and sampling diagnostics are descriptive. Rare categories must not be oversampled and presented as new independent coverage, and translated variants remain one template family for split and evaluation purposes.

Validation and limitations

Every emitted record is checked for identical schema/context, well-formed XML, valid nonoverlapping repair spans, no unlabelled changes, and exact recovery of clean values. The Parquet readback audit independently repeats these checks and verifies unique identifiers and disjoint source families. Contract tests cover valid formatting alternatives and category-specific partial repairs. These checks establish fidelity and tested contracts, not perfect natural-language ground truth.

Read audit.json before sampling: rare categories and category×language×split gaps must remain visible. Do not silently duplicate template translations across splits to fill gaps. This release provides training examples, not a trained detector, a calibrated error-rate estimate, or evidence that 8,192-token contexts are optimal.

Provenance and licensing

Included real data comes from USDA FoodData Central. USDA states that its data is public domain and published under CC0 1.0; see the FoodData Central API guide. Requested attribution: U.S. Department of Agriculture, Agricultural Research Service. FoodData Central, 2019. fdc.nal.usda.gov.

The corpus is assembled from open licensed data. Source-specific license identifiers, attribution requirements, and provenance are recorded in provenance.json; those terms remain authoritative for the corresponding records.

The combined dataset contains multiple source licenses. The project owns the synthetic data and licenses the published dataset under CC BY-NC 4.0. Commercial use requires permission from the project owner.