Blame for README.md

---
pretty_name: procedural-extraction
language:
- en
license: mit
task_categories:
- token-classification
- question-answering
tags:
- tasksource
- jev
- extraction
- clairic
- procedural
- synthetic
configs:
- config_name: record_field_extraction
  data_files:
  - split: train
    path: record_field_extraction/train-*.parquet
  - split: validation
    path: record_field_extraction/validation-*.parquet
  - split: test
    path: record_field_extraction/test-*.parquet
- config_name: state_update_extraction
  data_files:
  - split: train
    path: state_update_extraction/train-*.parquet
  - split: validation
    path: state_update_extraction/validation-*.parquet
  - split: test
    path: state_update_extraction/test-*.parquet
- config_name: conditional_list_extraction
  data_files:
  - split: train
    path: conditional_list_extraction/train-*.parquet
  - split: validation
    path: conditional_list_extraction/validation-*.parquet
  - split: test
    path: conditional_list_extraction/test-*.parquet
- config_name: log_query_extraction
  data_files:
  - split: train
    path: log_query_extraction/train-*.parquet
  - split: validation
    path: log_query_extraction/validation-*.parquet
  - split: test
    path: log_query_extraction/test-*.parquet
- config_name: typed_mention_extraction
  data_files:
  - split: train
    path: typed_mention_extraction/train-*.parquet
  - split: validation
    path: typed_mention_extraction/validation-*.parquet
  - split: test
    path: typed_mention_extraction/test-*.parquet
- config_name: business_document_extraction
  data_files:
  - split: train
    path: business_document_extraction/train-*.parquet
  - split: validation
    path: business_document_extraction/validation-*.parquet
  - split: test
    path: business_document_extraction/test-*.parquet
- config_name: nested_structure_extraction
  data_files:
  - split: train
    path: nested_structure_extraction/train-*.parquet
  - split: validation
    path: nested_structure_extraction/validation-*.parquet
  - split: test
    path: nested_structure_extraction/test-*.parquet
- config_name: html_extraction
  data_files:
  - split: train
    path: html_extraction/train-*.parquet
  - split: validation
    path: html_extraction/validation-*.parquet
  - split: test
    path: html_extraction/test-*.parquet
- config_name: date_range_extraction
  data_files:
  - split: train
    path: date_range_extraction/train-*.parquet
  - split: validation
    path: date_range_extraction/validation-*.parquet
  - split: test
    path: date_range_extraction/test-*.parquet
- config_name: spec_match_extraction
  data_files:
  - split: train
    path: spec_match_extraction/train-*.parquet
  - split: validation
    path: spec_match_extraction/validation-*.parquet
  - split: test
    path: spec_match_extraction/test-*.parquet
- config_name: lookalike_field_extraction
  data_files:
  - split: train
    path: lookalike_field_extraction/train-*.parquet
  - split: validation
    path: lookalike_field_extraction/validation-*.parquet
  - split: test
    path: lookalike_field_extraction/test-*.parquet
- config_name: structure_lookup_extraction
  data_files:
  - split: train
    path: structure_lookup_extraction/train-*.parquet
  - split: validation
    path: structure_lookup_extraction/validation-*.parquet
  - split: test
    path: structure_lookup_extraction/test-*.parquet
- config_name: relational_span_extraction
  data_files:
  - split: train
    path: relational_span_extraction/train-*.parquet
  - split: validation
    path: relational_span_extraction/validation-*.parquet
  - split: test
    path: relational_span_extraction/test-*.parquet
---

# procedural-extraction

Procedurally generated extraction requests: one text and several typed fields to fill from it in one pass, as
extraction models (e.g. Cloudflare's Clairic) are asked. Each field is `text` (one span) or `list` (every
matching span) and is answered with character spans of the text or null. Answers are computed by construction,
so spans are exact and labels noise-free (each generator samples a latent structure, such as a fact store,
a merged configuration tree or a DOM, and a solver computes the answer; independent readers of the rendered
texts recompute every answer in reasoning-core's tests); nulls are earned (the text mentions a negated, proposed, former or
other entity's value, not nothing). The tasks are generated by
[reasoning-core](https://github.com/sileod/reasoning-core) (commit `388d64a`, see
[provenance.json](provenance.json)) and are a procedural component of the `extraction` config of
[tasksource-jev-typed-decisions](https://huggingface.co/datasets/tasksource/tasksource-jev-typed-decisions).

| config | what is extracted |
|---|---|
| `record_field_extraction` | A person's city, employer, job title, phone, email and birth date, from prose about several people with near-miss names; negated, former, hoped-for and relatives' values as confusers; "same city/company as X" chains |
| `state_update_extraction` | The value in force (or before the last change) of plan slots, from a timestamped chat of statements, proposals approved/rejected/pending, corrections and undos, out of order from level 2 |
| `conditional_list_extraction` | Every order id or product meeting conjunctive, negated or numeric conditions after later updates, or the heaviest such order; prose, bullets or a table |
| `log_query_extraction` | Every request id or user of log lines matching filters on level, exact service (auth vs auth-proxy), status, latency and user, or the latest/slowest match; plain, logfmt or JSON lines |
| `typed_mention_extraction` | Named-entity style: every mention of each requested type, where one name (Jordan) is a person, place, organization or product by context; companies by a defined current role among former and unrelated relations; codes by a stated format among near-miss strings |
| `business_document_extraction` | The value in force on a date in an invoice, purchase order, contract or insurance policy: numbered amendments override printed values and each other, dated windows, schedule references that chain, asked keys left out while their near-miss key is stated (subtotal vs total) |
| `nested_structure_extraction` | Path, any-depth and predicate-filter queries on a JSON, YAML or XML configuration whose override layers deep-merge or replace whole blocks |
| `html_extraction` | Fields of the main item in raw HTML: scope (nested and related item cards), visibility (hidden, template, struck prices), dt/dd, label-for, attributes and spec tables with rowspan/colspan |
| `date_range_extraction` | Every invoice, shipment, contract or ticket whose date of one role (issued vs due) falls in a range, before/after a date, in a month or on a weekday, or the earliest/latest; dates ISO, worded, day/month or month/day numeric under a stated section convention, year-less under a stated year, or relative to another record |
| `spec_match_extraction` | Every laptop, e-bike or monitor meeting conjunctive spec requirements, or the best such one; specs and thresholds in mixed exact units (g/kg, GB/TB, min/h, Wh/kWh, in/cm) and a stated number convention (1.234,5), near-miss products failing one requirement, unstated specs not qualifying |
| `lookalike_field_extraction` | Fields asked by role (issue vs delivery date, subtotal vs tax, sender vs approver, order vs tracking number, billing vs return address, phones, weights, durations) from a form, sections, table, prose, email, log or YAML-like record; an asked role that is not stated is null although same-typed values of other roles are, or shows a placeholder (pending, N/A); a conditional or policy value of the role (a late fee that may apply) is not the answer |
| `structure_lookup_extraction` | A value by its place: a section path (Billing > Contact > Phone) in an outline whose labels recur under other sections, or a cell of one of several same-column tables picked by row and column, a condition on another column or a row position; Markdown, ASCII, CSV, aligned or transposed tables; null when the place exists only elsewhere or the cell is empty |
| `relational_span_extraction` | Every mention of an event's direct or indirect causes or effects in a narrative with temporal-only and ruled-out look-alikes, or of what follows from facts and if-then rules (some conjunctive), what triggers an event and what it triggers |

Levels 0 to 4 are drawn uniformly; `level` is a column. Mean per-field score by level (0 / 1 / 2 / 3 / 4):

| config | train | validation | test | DeepSeek-v4-flash | GLiNER-large |
|---|---|---|---|---|---|
| `record_field_extraction` | 10,000 | 500 | 500 | 1.00 / 0.96 / 0.81 / 0.72 / 0.62 | 0.80 / 0.59 / 0.36 / 0.45 / 0.38 |
| `state_update_extraction` | 10,000 | 500 | 500 | 0.84 / 0.76 / 0.77 / 0.68 / 0.51 | 0.58 / 0.46 / 0.37 / 0.36 / 0.29 |
| `conditional_list_extraction` | 10,000 | 500 | 500 | 0.88 / 0.79 / 0.70 / 0.64 / 0.52 | 0.29 / 0.17 / 0.12 / 0.05 / 0.10 |
| `log_query_extraction` | 10,000 | 500 | 500 | 0.86 / 0.76 / 0.68 / 0.63 / 0.59 | 0.31 / 0.16 / 0.09 / 0.24 / 0.15 |
| `typed_mention_extraction` | 10,000 | 500 | 500 | 0.85 / 0.80 / 0.79 / 0.69 / 0.66 | 0.40 / 0.42 / 0.49 / 0.34 / 0.39 |
| `business_document_extraction` | 10,000 | 500 | 500 | 0.70 / 0.77 / 0.76 / 0.81 / 0.75 | 0.53 / 0.49 / 0.45 / 0.40 / 0.30 |
| `nested_structure_extraction` | 10,000 | 500 | 500 | 0.93 / 0.88 / 0.94 / 0.86 / 0.86 | 0.18 / 0.20 / 0.20 / 0.20 / 0.18 |
| `html_extraction` | 10,000 | 500 | 500 | 0.98 / 0.96 / 0.94 / 0.98 / 0.90 | 0.35 / 0.31 / 0.30 / 0.27 / 0.39 |
| `date_range_extraction` | 10,000 | 500 | 500 | 0.75 / 0.62 / 0.60 / 0.44 / 0.44 | 0.16 / 0.03 / 0.03 / 0.00 / 0.02 |
| `spec_match_extraction` | 10,000 | 500 | 500 | 0.81 / 0.53 / 0.27 / 0.25 / 0.17 | 0.11 / 0.08 / 0.04 / 0.03 / 0.04 |
| `lookalike_field_extraction` | 10,000 | 500 | 500 | 0.93 / 0.96 / 0.89 / 0.85 / 0.80 | 0.37 / 0.35 / 0.23 / 0.22 / 0.09 |
| `structure_lookup_extraction` | 10,000 | 500 | 500 | 1.00 / 0.99 / 0.97 / 0.97 / 0.91 | 0.27 / 0.12 / 0.07 / 0.07 / 0.09 |
| `relational_span_extraction` | 10,000 | 500 | 500 | 0.83 / 0.85 / 0.85 / 0.84 / 0.73 | 0.16 / 0.11 / 0.09 / 0.07 / 0.07 |

## Schema

`state` (the text), `fields` (`[{name, type, instruction}]`), `answers` (one per field: `{null, values:
[{text, start, end}]}`, `state[start:end] == text`), `metadata` (JSON: level, state seed), `level`.

## 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). The procedural data is generated with Reasoning Core; please also cite the Reasoning Core paper (Sileo et al., 2026).

```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",
}

@article{sileo2026reasoning,
  title={Reasoning Core: Designing Broad Procedural Data for Completion-Supervised Reasoning Training},
  author={Sileo, Damien and Lacombe, Valentin and Kachler, Dimitri},
  journal={arXiv preprint arXiv:2608.05148},
  year={2026},
  url={https://arxiv.org/abs/2608.05148}
}
```