synthetic-typed-decisions

LLM-written items (emails, chats, forum threads, field reports, case notes, reviews...) with several typed questions over the same item, following the Jev / System One request shape: choice, noul (yes/no probability), and score (an ordered rubric). It covers judgment that cannot be computed by rules: meaning, intent, tone, stance, plausibility, urgency, fuzzy categorization and routing. Anything formalizable (rules, thresholds, deadlines, counting, lookups) is left to procedural-typed-decisions, whose labels are exact.

This is an independent dataset. It is not an official TypeSafe Jev dataset and is not produced by or affiliated with TypeSafe or OpenJev.

How it is made

1. Workflows. For a domain and a kind of incoming item, DeepSeek V4 Flash designs a reusable decision application: 3 to 8 short judgment questions over 32 skills, with natural option counts (2 to 8, and 12 to 40 for one routing or categorization question in some workflows).
2. Items. About 30 items are written per workflow. One or two questions per item get a sampled intended reading (a clear option, a borderline one, or a yes/no probability), so answers are balanced and some uncertainty is deliberate; the other questions are answered by whatever the item says.
3. Check. A separate blind DeepSeek pass answers every question with probabilities and flags ill-posed ones, which are dropped.
4. Labels. Jev 1.13 (typesafe/jev-1.13 on OpenRouter) labels every kept question; questions where Jev and the check confidently disagree are dropped. answers averages Jev's probabilities with the blind check's. On 75 questions labelled blind by an independent annotator (Claude), the average was closer than either alone (total variation 0.135 vs 0.192 for Jev and 0.179 for the check; same top answer 87% vs 83% and 84%). teachers holds each model's probabilities, including upstage/solar-decide (up to 26 options; 0.304) and, for yes/no questions, respan/span-01.

The soft labels carry real uncertainty (see the confidence shares below); train on the full distributions rather than the argmax.

Fields

field	content
state	the item text
questions	JSON Jev request: {id: {type, instructions, criteria}}
answers	JSON answers in Jev format with the averaged probabilities (noul, or probabilities over the criteria)
teachers	JSON {id: {model: probabilities}} for each decision model
skills	JSON {id: skill}
workflow, application, domain, source	the decision application; domain is a loose tag

Size and splits

Splits are by workflow, so validation and test use decision applications never seen in training.

split	items	decisions	workflows
train	6,666	21,190	235
validation	1,176	3,996	41
test	1,074	3,475	38

Decisions: 15,585 choice, 7,762 noul, 5,314 score. The labels are at least 0.95 confident on 11% of them and below 0.7 on 44%.

Caveats

Items are synthetic and the labels come from a model. A few questions still lean on an unstated standard (for example "standard safety procedures"). The generator code is src/tasksource/jev/synthetic/workflows.py in tasksource.

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)",
    month = may,
    year = "2024",
    address = "Torino, Italia",
    publisher = "ELRA and ICCL",
    url = "https://aclanthology.org/2024.lrec-main.1361",
    pages = "15655--15684",
}