Datasets:

HiTZ
/

critics_science_mt_benchmark

# Canonical JSONL schema v1.1.0

One JSON object per source excerpt, one line per item. Full specification: `schema/benchmark.schema.json` (JSON Schema Draft 2020-12).

## Primary key and grouping

- `id` is the original, stable excerpt key, globally unique across batches. Duplicate IDs across batches are rejected even if the text matches.
- `batch_id` identifies the originating input folder, `batch_001` to `batch_010`; it is NOT a train/test split.
- `source.source_identifier` retains the original source ID.
- `source.source_document_key` is a **provisional** document grouping key derived from the source ID, not guaranteed to be an authoritative article identifier. `source_document_key_status` records this limitation.
- `source.article_doi`, `article_url`, `license`, `excerpt_location` are nullable pending source verification; never populate these with guesses.
- `source.source_sha256` hashes the UTF-8 content of the exact English text (no normalization).

## Translations

- `translations.es.mt` and `translations.eu.mt` hold the exact machine translation and its supplied metadata. Do not alter or re-label the original generation model.
- `translations.es.post_edit` and `translations.eu.post_edit` are `null` until real human post-edits exist. Never copy MT text into these fields as placeholder references.
- A future post-edit object requires `text`, `status` (`draft`, `reviewed`, `adjudicated`), `reviewed_date`, `reviewer_id` (pseudonymous), `revision`; optional `notes`.

## Evaluation-only flattening

The builder generates `data/pairs/en-es/test.jsonl` and `data/pairs/en-eu/test.jsonl`, with fields `id`, `batch_id`, `topic`, `repository`, `source_document_key`, `source_text`, `source_sha256`, `target_lang`, `mt_text`, `reference_text`, `reference_status`, `mt_system`, `mt_prompt_version`.

There is deliberately **no training split** and no fine-tuning export. Candidate MT and reference are separate fields, so evaluators can elect to distribute only input text and score submitted systems using withheld references.

## Determinism and invariants

Builder rejects missing files, within-file duplicate IDs, misaligned language ID sets, missing source/target text, malformed translation metadata, duplicate IDs across all loaded batches, invalid schema records and unrecognized topics.

The schema cannot establish translation quality, article provenance or legal redistribution rights. Those remain manual-review requirements.

Optional human corrections are provided separately as `post_edits/batch_NNN/es.json` and `eu.json` using an array of records with `id` and fields required by the post-edit schema.

## Storage paths

The canonical multilingual file is `data/canonical/all_languages.jsonl`. Generated per-batch canonical files are in `data/canonical/batch_NNN.jsonl`; the pair views are in `data/pairs/` and batch validation output is in `data/reports/`. Raw input files stay unchanged in `original_input/`.

	# Canonical JSONL schema v1.1.0

	One JSON object per source excerpt, one line per item. Full specification: `schema/benchmark.schema.json` (JSON Schema Draft 2020-12).

	## Primary key and grouping

	- `id` is the original, stable excerpt key, globally unique across batches. Duplicate IDs across batches are rejected even if the text matches.
	- `batch_id` identifies the originating input folder, `batch_001` to `batch_010`; it is NOT a train/test split.
	- `source.source_identifier` retains the original source ID.
	- `source.source_document_key` is a provisional document grouping key derived from the source ID, not guaranteed to be an authoritative article identifier. `source_document_key_status` records this limitation.
	- `source.article_doi`, `article_url`, `license`, `excerpt_location` are nullable pending source verification; never populate these with guesses.
	- `source.source_sha256` hashes the UTF-8 content of the exact English text (no normalization).

	## Translations

	- `translations.es.mt` and `translations.eu.mt` hold the exact machine translation and its supplied metadata. Do not alter or re-label the original generation model.
	- `translations.es.post_edit` and `translations.eu.post_edit` are `null` until real human post-edits exist. Never copy MT text into these fields as placeholder references.
	- A future post-edit object requires `text`, `status` (`draft`, `reviewed`, `adjudicated`), `reviewed_date`, `reviewer_id` (pseudonymous), `revision`; optional `notes`.

	## Evaluation-only flattening

	The builder generates `data/pairs/en-es/test.jsonl` and `data/pairs/en-eu/test.jsonl`, with fields `id`, `batch_id`, `topic`, `repository`, `source_document_key`, `source_text`, `source_sha256`, `target_lang`, `mt_text`, `reference_text`, `reference_status`, `mt_system`, `mt_prompt_version`.

	There is deliberately no training split and no fine-tuning export. Candidate MT and reference are separate fields, so evaluators can elect to distribute only input text and score submitted systems using withheld references.

	## Determinism and invariants

	Builder rejects missing files, within-file duplicate IDs, misaligned language ID sets, missing source/target text, malformed translation metadata, duplicate IDs across all loaded batches, invalid schema records and unrecognized topics.

	The schema cannot establish translation quality, article provenance or legal redistribution rights. Those remain manual-review requirements.

	Optional human corrections are provided separately as `post_edits/batch_NNN/es.json` and `eu.json` using an array of records with `id` and fields required by the post-edit schema.

	## Storage paths

	The canonical multilingual file is `data/canonical/all_languages.jsonl`. Generated per-batch canonical files are in `data/canonical/batch_NNN.jsonl`; the pair views are in `data/pairs/` and batch validation output is in `data/reports/`. Raw input files stay unchanged in `original_input/`.