## Extraction config

`load_dataset("tasksource/tasksource-jev-typed-decisions", "extraction")`: one text and several typed fields to fill from it, answered in one
pass, as extraction models (e.g. Cloudflare's Clairic) are asked. Rows: 552,917 train, 44,205 validation, 45,112 test.

- `state`: the text. `fields`: `[{name, type, instruction}]`; `type` is `text` (one span; the instruction is
  a question) or `list` (every span of a kind, e.g. every person).
- `answers`: one per field, in order: `{null, values: [{text, start, end}]}`, character offsets into
  `state` (`state[start:end] == text`, checked for every span). `null` means the text does not hold the
  answer (an unanswerable question, an absent entity type). Several values of a `text` field are
  alternative gold spans; of a `list` field, every item.
- Extractive question answering (SQuAD 2.0, AdversarialQA, Natural Questions, HotpotQA, Quoref, MuSiQue,
  DROP single-span, DuoRC, ConcurrentQA, FairytaleQA, QuAC with history, windowed CUAD, TyDi English,
  Qasper spans, PolicyQA, SubjQA, COVID-QA, TimeQA, TAT-QA text spans, ROPES, QAConv and MultiDoc2Dial
  grounding, MultiSpanQA, QA-SRL) groups the questions asked about one context, up to 8 fields per row.
  Named entities (CoNLL-2003, OntoNotes, WNUT-17, Few-NERD, NCBI-disease, JNLPBA, acronyms, FiNER-139,
  MAVEN event triggers, SLURP slots, EBM-NLP PICO spans, and astronomy, biomedical, financial, software,
  legal, GUM, Twitter and English opinion-target sources) ask for every entity type of the dataset, or
  the types present plus absent ones up to 8. Entity texts are the source tokens joined by spaces.
  Other span tagging (skills, keyphrases, aspect terms, toxic spans, metaphors, learner errors, BioScope and
  ConanDoyle-neg negation and speculation scopes, AbstRCT claims and premises, GUM elementary discourse units)
  asks for every span of its kind. Event arguments (PHEE adverse drug events, RAMS and CASIE) ask for every
  argument of each role of a trigger given by its character offsets.
- Structural extraction uses English UD EWT/GUM dependency subtrees (`ud/...`), PreCo coreference
  mention sets (`preco/mentions`), human-revised Re-DocRED relation targets (`redocred/entities`),
  and CoNLL-2000 phrase chunks. UD fields select a unique head and keep full contiguous subtrees;
  a missing dependent is null and a non-projective positive field is omitted.
- Procedural extraction ([procedural-extraction](https://huggingface.co/datasets/tasksource/procedural-extraction),
  sources `procedural/...`): person records with near-miss names and confusers, plan slots in force after a chat
  of proposals and undos, conditional order lists after updates, log queries, typed mentions, business documents
  with amendments, nested configurations with override layers, raw HTML pages, records dated in mixed formats
  within a range and products matching specs in mixed units and number styles, at difficulty levels 0 to 4;
  spans and nulls exact by construction.
- Mostly English, with a small multilingual share: TyDi QA (Arabic, Bengali, Finnish, Korean, Russian, Swahili),
  UNER (Portuguese, Chinese), MasakhaNER 2 (Swahili) and MASSIVE slots (Hindi), at most 1,600 training rows per
  language.
- Native splits, except those without released labels; evaluation texts that also occur in training are dropped. Each source is capped at
  10,000 training rows (a suite of configs, such as `swde/...`, shares 20,000) and 500 per evaluation split; at most 20% of an entity source's
  texts mention none of its entity types. `license`/`license_use` per row;
  `metadata` keeps source question ids or entity types. See [build-manifest.json](build-manifest.json).

The same question answering sources load in SQuAD format for `AutoModelForQuestionAnswering` with
`tasksource.load_task(id, extractive=True)`.