This dataset contains 60,365 decision records, mixed from three stages and shuffled with seed 20261014. It is intended for supervised fine-tuning of structured decision models. It contains a single training split; no held-out evaluation split has been created.

Tasks, formats, and size

Property	Value
Tasks	Text Classification; Multiple Choice; Question Answering
Subtasks	Multi-class classification; intent classification; natural language inference; multiple-choice QA
Formats	JSON Lines (`train.jsonl`) and JSON array (`train.json`); JSON format on the Hub
Size	60,365 records; `10K<n<100K`
`train.jsonl` file size	39,375,420 bytes (39.38 MB)
`train.json` file size	48,117,723 bytes (48.12 MB)
Languages	Primarily English; one fallback option key is Vietnamese

The YAML metadata declares the Hub tasks, size category, formats, and decision-data tags. decision-data is a descriptive tag; the official task categories are text classification, multiple choice, and question answering.

Files

train.jsonl: one decision record per line; use this file with kev.train --data.
train.json: the same records in the same order, formatted as a readable JSON array.
manifest.json: public output checksums, aggregate counts, schema information, and shuffle settings.

The JSON and JSONL files are two serializations of the same dataset. Load only one of them. The native schema has boolean and string labels in different question types; read it with Python's json module or the Kev loader rather than relying on automatic Arrow schema inference.

Composition

Input	Records
Stage 1	35,785
Stage 2 minimal pairs	17,580
Stage 3	7,000
Total	60,365

Stage 2 contributes both members of each minimal pair: the positive record includes the correct option plus NOTA, and the negative record removes the correct option and labels NOTA. stage2_30pct.json is not added separately. After global shuffling, pair members need not be adjacent.

Data types

The mixture covers general multiple-choice reasoning and science QA, reading comprehension, natural-language inference, binary yes/no decisions, financial-product classification, contract-clause classification, and banking-intent routing. Stage 2 adds mechanically generated positive/negative minimal pairs for eligible multiple-choice records.

Schema

{"id":"99ba76d4-79a5-43bc-9e18-302be4e7201b","state":"Document or question context","questions":{"answer":{"type":"choice","instructions":"Choose the correct answer.","criteria":{"A":"First option","B":"Second option"},"label":"A"}}}

Each record contains exactly id, state, and questions. Each question contains its type, instructions, label, and criteria where applicable:

choice: the string label exactly matches a criteria key; a criteria value may be a description or null.
noul: the label is a JSON boolean.

All instructions are in English. Original text and option names are preserved; one financial-product fallback option uses Khác (Other). Question IDs vary by task (answer, product, clause_type, intent).

Record IDs

Each record has a random UUID4, for example 99ba76d4-79a5-43bc-9e18-302be4e7201b. IDs do not encode the dataset, stage, source row, or pair variant. Duplicate source rows and the two members of each minimal pair receive distinct IDs. Training records retain exactly the same IDs as their corresponding stage records; shuffling never renumbers them. The only additional record field is id. Rebuilding stages creates fresh UUIDs. Rebuilding only the training mix from unchanged stage files preserves all IDs and the seeded shuffle order.

Preparation and reproducibility

Stage 1 contains converted decision tasks. Stage 2 contains minimal-pair variants of eligible multiple-choice records. Stage 3 adds complex document and domain-intent classification. The merge uses all input records without resampling or deduplication. Python random.Random(20261014).shuffle is applied once to the complete concatenated record list. There are 45,938 choice questions and 14,427 yes/no questions. Exact duplicate records are preserved (528 extra occurrences).

In the pipeline repository, rebuild with:

uv run decision-data build-training --output-dir data/data_decision --seed 20261014

Reading and training

import json

with open("train.jsonl", encoding="utf-8") as stream:
    for line in stream:
        record = json.loads(line)
        # record["state"], record["questions"]

In an installed Kev repository:

uv run python -m kev.train --data /path/to/train.jsonl --init_from /path/to/base-model --lr 2e-5 --out runs/decision-data-v1

For Hugging Face, upload the contents of this directory so that this README is at the dataset repository root. Use train.jsonl as the training artifact; train.json is an optional readable copy.

Upload from the pipeline repository

uv run hf auth login
uv run decision-data upload-training --repo-id Phanhai2804/decision-data-v1 --dry-run
uv run decision-data upload-training --repo-id Phanhai2804/decision-data-v1

Authentication uses the cached Hugging Face login or HF_TOKEN; no token belongs in the code or this README. Dry-run validates the export locally and makes no Hub requests. Upload includes only README.md, manifest.json, train.jsonl, and train.json, at the repository root. It does not delete remote files or change an existing repository's visibility.

The configs metadata selects only train.jsonl for the train split, so the readable JSON copy and manifest are not counted as extra training data. The nested native schema includes mixed string/boolean labels; generic Arrow-based previews may require a custom reader. Metadata does not guarantee successful automatic Dataset Viewer parsing.

Known limitations

The financial complaint subset retains 1,000 selected records: 729 use issue/sub-issue/company instead of a missing narrative; 26 labels outside the ten most common products are mapped to Khác. Identical fallback inputs have contradictory labels in 391 records. These records are retained, not silently corrected.
Parts of this collection may overlap with common benchmark validation/test material. Do not report evaluation on overlapping benchmarks as held out.
Split future evaluation data by original example before generating pairs, and keep duplicates and both members of a minimal pair in the same split to avoid leakage.
Mechanical minimal pairs test the removal of the correct option; they do not constitute a complete semantic data-quality review.

Usage and provenance

This is a transformed multi-domain collection. No single permissive license is asserted for the entire mixture. Users are responsible for ensuring that their intended use and redistribution comply with all applicable terms. Detailed provenance is intentionally not included in this public card.

Output checksums

train.jsonl: 9f7539920aa7f44eb9cd8fe783a1ed75c6abbbd8e2de8705d113727efd92879f
train.json: 34ab55f2fc0169c06a2e3180c848812dab9ef644966d12bd3acf96575cd75abc

Downloads last month: 26

Total file size:

87.5 MB