tasksource-jev-typed-decisions

2.5 million typed decisions (choices, ratings and probabilities) from 670 sources.

Why use it

• Real supervision. Labels, ratings, and annotator votes come from established datasets, not a teacher model. Every row names its source.
• Breadth. Over 300 dataset families: NLI and reasoning, QA and commonsense, sentiment, intent and topic, toxicity and safety, preference pairs, fact checking, entity tagging, and dozens of languages. GLUE, SuperGLUE, HellaSwag, PIQA, ScienceQA, Banking77, CoNLL-2003, MasakhaNEWS, HelpSteer, ChaosNLI, and many more, with no task allowed to dominate.
• Three decision types in one schema. choice (pick one option), score (an ordered scale), and noul (the probability that the answer to a yes/no question is yes). noul holds only probabilities: entailment likelihoods, and the share of annotators who answered yes. Mean ratings and similarity are score distributions whose expected level is the mean (3.4 on 1–5 puts 0.6 on 3 and 0.4 on 4). Ordinal label sets appear as both choice and score, split deterministically per row, so a model learns both requests for the same scale. Soft targets are kept wherever the source has mean ratings or votes from at least five annotators per item (vote shares from fewer are too noisy): STS, ChaosNLI, civil_comments, Measuring Hate Speech, WouldYouRather, ProtoQA, LeWiDi and more. They make up the graded share.
• Built so position, repeated eval data, and question choice give nothing away.
  • Multiple-choice options are shuffled per row, so the answer's position carries no signal.
  • Validation and test rows whose content appears in train are removed.
  • Derived questions are chosen without looking at their answers.
  • Annotations were reviewed task by task. Inverted, unanswerable, and garbled labels were fixed or dropped.
• Multi-question states. Related decisions share a group_id and can be asked together. Packed states test reasoning over several items at once, and procedural-typed-decisions adds exact counting, arithmetic, retrieval, state tracking, routing among up to 60 options, and exact posteriors when a policy applies to a requester whose role is uncertain.

Quick start

Coding agent? Read AGENTS.md: row semantics, rebuilding multi-question requests from group_id, filtering, and evaluation caveats.

from datasets import load_dataset

ds = load_dataset("tasksource/tasksource-jev-typed-decisions")          # steered 1M-row mix
# full = load_dataset("tasksource/tasksource-jev-typed-decisions", "full")  # every row of the build
row = ds["train"][0]
print(row["state"], row["question"], row["options"], row["target"])

{"state": "My body cast a shadow over the grass. What was the cause of this?",
 "question": "Choose the criterion that best answers the question.",
 "kind": "choice", "options": ["The sun was rising.", "The grass was cut."],
 "target": [1.0, 0.0], "source": "super_glue/copa"}

Configs

• default: a steered mix of about 1M train rows. Sources are first gated on label correctness, then weighted by how interesting they are and how close they sit to the zone of proximal development (judged by decision models). Two-option tasks get fewer rows, and procedural generators get 12%. No row is repeated. Validation and test are the full eval splits, restricted to the mixed sources. Buckets and shares are in jev_mixes.py and per-source scores in jev_source_scores.csv.
• full: every row that passed the build, with per-source caps (about 2.5M train rows).

Format

field	meaning
state	The text to decide about
question	What to decide
kind	choice, score, or noul
options	Runtime criteria; empty for noul
target	Distribution over options, or [p] for noul
id, group_id, question_id	Link decisions over the same source example
source, split, variant	Originating task, original split, and recast variant
license, license_use	The source's license(s), and commercial, non-commercial or unspecified (see below)

Splits: train (about 1M rows in default, 2.5M in full), 15,000 validation (dev in split), and 15,000 test, following each source's own train/dev/test splits where it has them.

How it is built

• Canonical recasts. Each Tasksource task is converted deterministically.
  • Criteria are the source's own label names and answer options.
  • Multiple-choice rows keep every option in a per-row order.
  • A final "all/none of the above" reads "all/none of the other options".
  • Options that cite other options by letter or number keep their order.
  • The question is the task's own when its inputs alone do not say what to predict ("What stance does the tweet take on feminism?"), and a generic instruction otherwise. Label-verification and packed questions carry it too.
• Variants. Low-frequency, deterministic variants cover label verification as noul, criterion order, and instruction wording.
• Packing. Up to 10% of each classification task's examples are packed, two to four at a time, into packed_derived states. Their questions (an item's label, agreement, existence, counts) follow exactly from the gold labels.
• Mixing (full). Formats get fixed shares of the train rows (47% classification, 30% multiple choice, 3% token labeling, 10% graded (soft-label sources), 10% procedural). Within a format, dataset families get equal shares, scaled by hand-set weights (more for adversarial NLI, long documents and preference pairs; less for templated probes), times audit weights from a per-task check of Jev on 200 examples: ×1.5 for hard tasks whose gold is right by construction (synthetic logic, theory of mind, spatial reasoning), ×0.5 for near-solved tasks and for hard tasks whose gold is a judgment call (ratings, preferences, crowd sentiment). Sources with many options get slightly more room. Related questions are kept together.
• Mixing (default). About 1M rows drawn from full, never repeating a row. Buckets of related sources get set shares (jev_mixes.py: 15% logic, 10% NLI, 10% knowledge QA, 9% long documents and fact-checking, 8% intent and routing, 12% procedural, ...). Inside a bucket, sources get rows by score × √size. The score multiplies:
  • correctness: sources that failed label review get 0;
  • the zone of proximal development: how much probability the decision models (Jev, Liquid D1) give the gold answer. Near-solved sources and sources the models miss outright both get less;
  • interest and cleanliness, from Jev yes/no checks for transferable skills, trivial examples and malformed rows;
  • ×0.6 for two-option tasks (about a third of the rows);
  • ×2 for sources picked by reading them.
• Order and coverage.
  • The first 1,000 train rows are interleaved to show variety in the Dataset Viewer; the rest is shuffled. Questions of a group stay adjacent throughout.
  • Evaluation benchmarks (BIG-bench, MMLU, BLiMP, MATH test, ...) are left out so they stay clean for evaluation.
• Sources. sources.yaml lists every source with its rows, the Hub dataset and revision it was loaded from, the original dataset behind each tasksource copy, and its licenses.
• Audit trail. The source mix, failed source list, and build manifest ship with the data.
• Reproducible. The build runbook rebuilds the release from Tasksource's task catalog.

License and scope

Tasksource harmonizes datasets from many publishers; their original licenses and terms still apply, hence license: other.

Each row carries its source's license, to help filter:

ds = ds.filter(lambda use: use == "commercial", input_columns="license_use")

• license lists the license of the Hub dataset card the source was loaded from, and of the original dataset behind a tasksource copy. It also lists licenses recorded by the Data Provenance Initiative, marked (DPI).
• license_use takes the most restrictive of those: non-commercial if any is non-commercial or academic-only, commercial if one allows commercial use (share-alike and copyleft included), and unspecified otherwise. That covers missing licenses and other, bare cc, and no-derivatives licenses.
• sources.yaml records each card and DPI license per source.

This is a best-effort aid, not legal advice. Licenses on cards can be wrong or incomplete, and a source's terms may differ from its card's. Check the original terms before relying on them. This recast is independent of TypeSafe and OpenJev.

Citation

@inproceedings{sileo-2024-tasksource,
  title = {tasksource: A Large Collection of {NLP} tasks with a Structured Dataset Preprocessing Framework},
  author = {Sileo, Damien},
  booktitle = {Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
  year = {2024},
  pages = {15655--15684},
  url = {https://aclanthology.org/2024.lrec-main.1361/}
}