DATA_PROVENANCE.md

# Data provenance

This document is the normative source-to-release record for
`bactrianus/bactrainus-hotpotqa`.

## Upstream identity

| Property | Value |
|---|---|
| Repository | `hotpotqa/hotpot_qa` |
| Immutable revision | `1908d6afbbead072334abe2965f91bd2709910ab` |
| Configuration | `distractor` |
| Split | `train` |
| Expected rows | 90,447 |
| Candidate paragraphs per row | 2-10; 89,609 rows contain 10 |
| License | CC BY-SA 4.0 |

The official fields are `id`, `question`, `answer`, `type`, `level`, `context`,
and `supporting_facts`. Construction is pinned to the revision above and never
joins data by DataFrame index, filename order, or directory enumeration.

## Canonical normalization

For each upstream object with identifier \(i\), the builder emits one
`structured` record with `source_id = i`:

- `id` becomes `source_id`;
- `type` becomes `question_type`;
- `level` becomes `difficulty`;
- each context title and sentence list becomes one `candidate_paragraphs` item;
- each supporting title and index becomes one `supporting_facts` item; and
- `gold_paragraph_titles` is the order-preserving set of referenced titles.

For context \(C_i\) and supporting facts \(S_i\), every record must satisfy

\[
2\le |C_i|\le10,\qquad
\forall(t,j)\in S_i:\;t\in\operatorname{titles}(C_i)
\land 0\le j<|C_i[t]|.
\]

## ID-aligned task views

All SFT records retain `source_id` and are pure functions of the canonical
record.

| Configuration | Transformation |
|---|---|
| `reader-sft` | Gold evidence to short answer |
| `cot-reader-sft` | Gold evidence to an indexed evidence trace and answer |
| `paragraph-selector-sft` | Complete candidate set to supporting paragraph titles |
| `question-decomposer-sft` | Selected paragraphs to ordered grounded sub-questions |
| `sentence-selector-sft` | Selected paragraphs to supporting title/index pairs |
| `decomposed-sentence-selector-sft` | Selected paragraphs plus sub-questions to supporting title/index pairs |
| `joint-selector-reader-sft` | Complete candidate set to supporting facts and answer |

The evidence trace copies the annotated sentences exactly. The decomposition is
constructed from the paragraphs supplied to the decomposer. Both transformations
are deterministic, complete, and covered by the same cross-configuration ID
checks as the other SFT views.

## Build and validation chain

```bash
python scripts/build_release.py --root .
python scripts/validate_release.py --root .
python scripts/publish_release.py --root .
```

The builder uses private staging and refuses to overwrite an existing release.
The validator checks schema, row count, unique IDs, cross-view ID equality,
paragraph count, evidence bounds, difficulty totals, chat roles, and task
identifiers. Only then does it write `CHECKSUMS.sha256`.

`SOURCE_MANIFEST.json` is machine-readable and records the source revision and
all published configurations. A local official export may be used instead; in
that case the manifest records its filename, byte size, and SHA-256 digest.

The pinned source contains 22 out-of-range supporting-fact indices. The complete
reviewed mapping is versioned in `SOURCE_PATCHES.json`. Repairs are keyed by
`source_id`, title, and original index; they either select the in-range sentence
in the same paragraph that states the missing relation or remove a redundant
invalid fact when valid annotations in that paragraph already carry the
evidence. The original question, answer, context text, and all 90,447 IDs remain
unchanged.

## Published layout

```text
data/
|-- structured/train-*.parquet
|-- reader-sft/train-*.parquet
|-- cot-reader-sft/train-*.parquet
|-- paragraph-selector-sft/train-*.parquet
|-- question-decomposer-sft/train-*.parquet
|-- sentence-selector-sft/train-*.parquet
|-- decomposed-sentence-selector-sft/train-*.parquet
`-- joint-selector-reader-sft/train-*.parquet
```

Every content change requires regenerated Parquet shards, checksums, a new
dataset revision, and a recorded immutable commit SHA for downstream use.