# Changelog

Notable changes to this dataset.

## 2026-08-16

**No data changed in this revision.** Every row, every field and every value is what
the loading script produced. What changed is that the data is now in the repository
instead of being fetched from elsewhere at load time.

### Fixed

- **`load_dataset("iapp/iapp_wiki_qa_squad")` works again.** It had been failing since
  `datasets` dropped support for loading scripts at v3. This repository held a script
  and no data at all: `iapp_wiki_qa_squad.py` downloaded `squad_format/data.zip` from
  GitHub every time anyone loaded the dataset. With script support gone, `load_dataset`
  failed and every datasets-server endpoint returned HTTP 500, so the dataset viewer,
  preview, search, filter and statistics were all dead.

  The consequence was that our own published dataset was reachable in practice only
  through third-party re-uploads that we do not control and whose parity with the
  canonical data nobody had confirmed.

### Added

- `data/train-00000-of-00001.parquet`, `data/validation-00000-of-00001.parquet` and
  `data/test-00000-of-00001.parquet`, built from `squad_format/data.zip` in
  `iapp-technology/iapp-wiki-qa-dataset`, which remains the canonical source.

  These were checked against a run of the old loading script on `datasets` 2.x before
  publication: same six fields in the same order, same types, same rows in the same
  order, all three splits, **byte-for-byte identical**. That check is the whole basis on
  which the script was allowed to be deleted.

- `LICENSE`, MIT, carrying iApp's copyright line for the annotation work, plus a note
  that the MIT grant does not reach the `context` column. That column is Thai Wikipedia
  article text under CC BY-SA, and share-alike attaches to it in any redistribution
  regardless of what we license our own contribution as. The card previously declared
  `license: mit` while linking to a `LICENSE` file that returned 404.

- `validate.py`, which fetches the canonical archive, applies the six-field projection
  itself, and compares the result against the published parquet row by row. It also
  checks that every answer span is exactly the slice of context it claims to be, that
  no article appears in more than one split, and that the three withheld fields below
  are absent. 23 checks, all passing at this revision, non-zero exit on any failure.

- A `configs` block naming the config `iapp_wiki_qa_squad`, which is what the loading
  script's single `BuilderConfig` was called, so code that passes that name explicitly
  keeps working.

### Removed

- `iapp_wiki_qa_squad.py`, the loading script.

### Note on three fields the source carries and this dataset does not

The records in `squad_format/data.zip` have nine fields. The loading script emitted
six. The three it dropped are `created_by`, an account identifier resolving to 57
individual annotators; `created_on`; and `is_pay`, which holds the payment status,
date and amount for each annotated item.

Nothing recorded that the projection was doing that job, and materialising the archive
without reproducing it would have published annotator identities and their per-item pay
into a dataset viewer that indexes, searches and filters. The projection is therefore
reproduced exactly rather than reinvented, and `validate.py` fails if those fields ever
appear.

The worked example printed in earlier revisions of the dataset card included all three,
with a real account identifier and a real payment record, even though the loaded
dataset never contained them. That example has been replaced with what `load_dataset`
actually returns.

### Unchanged

- Row counts: 5,761 train, 742 validation, 739 test, from 1,529, 191 and 192 articles.
- Row order, field names, field order, field types, split names. The `validation` split
  is still called `validation`, not `valid`, as it was under the loading script.
- A score computed on this dataset before this revision is comparable with one computed
  after it.