---
pretty_name: synthetic-extraction
language:
- en
license: apache-2.0
task_categories:
- token-classification
- question-answering
tags:
- tasksource
- jev
- extraction
- clairic
- synthetic
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
  - split: validation
    path: data/validation-*.parquet
  - split: test
    path: data/test-*.parquet
---

# synthetic-extraction

LLM-written extraction requests: one natural document and the fields an application would extract from it, typed
`text` (one span) or `list` (every matching span), answered with character spans or null, as extraction models
(e.g. Cloudflare's Clairic) are asked. It covers fields that need reading rather than a solver (the corrected date in
a thread, who is assigned an action, the item a complaint is about); computable fields are in
[procedural-extraction](https://huggingface.co/datasets/tasksource/procedural-extraction). It is a small booster for
the `extraction` config of
[tasksource-jev-typed-decisions](https://huggingface.co/datasets/tasksource/tasksource-jev-typed-decisions).

## How it is made

1. **Design**: per domain and document source, `deepseek-v4-flash-0731` drafts an application's schema: a document kind
   and typed fields with precise instructions.
2. **Write**: it writes documents with an intended answer per field, each planned as present, present beside a near
   miss (a similar value that does not qualify), or absent beside a near miss, so nulls are earned.
3. **Verify**: answers must be verbatim spans. Two blind readers, one of another model family
   (`gpt-oss-120b`, `deepseek-v4-flash-0731`) fill every field. A field is kept when both agree with the
   writer (up to case, edge punctuation and a leading article); when one disagrees, a judge sees both answers in a
   random order and the field is kept only for an unambiguous verdict for the writer. When both disagree it is
   dropped, unless both copy the same span and the judge rules for it over the writer's answer (the writer missed a
   value it wrote). Documents keep their verified fields (at least 2).

43,189 of 53,090 written fields passed verification. Splits are by schema: validation and
test schemas never occur in train. Prompts, models and the generation commit are in
[provenance.json](provenance.json).

| split | rows |
|---|---|
| train | 7,670 |
| validation | 1,035 |
| test | 1,124 |

## Schema

`state` (the document), `fields` (`[{name, type, instruction}]`), `answers` (one per field: `{null, values:
[{text, start, end}]}`, `state[start:end] == text`), `metadata` (JSON: schema, domain, document source, field
plans, judged and dropped fields).

## Citation

If you use tasksource, its standardized task annotations, or datasets derived from them in published research, please cite the tasksource paper (Sileo, LREC-COLING 2024).

```bibtex
@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",
    doi = "10.63317/2h4pkr685mkq",
    pages = "15655--15684",
}
```