---
language:
- fr
- de
task_categories:
- token-classification
tags:
- historical-newspapers
- named-entity-recognition
- impresso
pretty_name: Impresso Media Sources Dataset
license: other
---

# Impresso Media Sources Dataset

Curated token-classification data for news-agency and radio-station mentions in Impresso historical newspaper text.

The v0.1 data is derived from the legacy French/German HIPE-style news-agency annotations, converted to JSONL, manually reviewed against the current model's dev/test disagreements, and updated according to annotation guidelines v2.0. The current guidelines annotate every explicit canonical media-source organization mention, not only source-attribution uses.

## Files

```text
data/train.jsonl
data/validation.jsonl
data/test.jsonl
label_map.json
DATASET_STATISTICS.md
```

`DATASET_STATISTICS.md` is a generated, human-readable release report with split sizes, token and mention totals, language and entity-family distributions, date and newspaper coverage, per-label frequencies, and the train/validation/test document-overlap check.

Each JSONL row represents one document/article. The published `data/*.jsonl` files intentionally use a compact training schema rather than the full converted HIPE payload, so they are easy to load with `datasets` and inspect in the Hub viewer:

- `id`, `language`, `newspaper`, `date`, `year`, `document_id`
- `text`
- `tokens`
- `token_start_offsets`, `token_end_offsets`
- `token_labels`
- `entities`
- `quality_flags`
- `legacy`

The canonical annotations are the BIO labels in `token_labels` plus the resolved span records in `entities`. The row-level `legacy` object is only for tracing back to the converted HIPE source and may contain `source_format` and `source_file`.

Important metadata fields:

- `newspaper`: source newspaper/media identifier from the original annotation metadata, for example `DTT`.
- `legacy.source_format`: original annotation format, for example `hipe-tsv`.
- `legacy.source_file`: original converted annotation file, for example `data/annotated_data/de/newsagency-data-dev-de.tsv`.
- `quality_flags`: non-fatal import or curation warnings. `has_forbidden_legacy_labels` means the original row contained labels excluded by the current label policy, such as `unk`, unresolved `ag`, or `pers.ind.articleauthor`; these labels were removed from the trainable annotation.
- `entities[].wikidata_url`: canonical Wikidata URL when available. Raw NEL/QID provenance is not part of the public training contract.
- `entities[].ocr_correction`: optional object present only when OCR/transcript evidence corrected the visible entity surface.
- Public `entities[]` do not include synthetic entity IDs. If a stable entity reference is needed, derive it from row `id`, character offsets, and `label`.

Fields intentionally excluded from the public training rows:

- `token_nel`, `token_ocr`, `token_render`, `token_segment_ids`: token-level HIPE side channels used for conversion/debugging, not model training.
- `token_label_ids`: integer labels are derived from `token_labels` and `label_map.json` by training code when needed.
- `segments`: segment and IIIF metadata; useful for traceability, too large for the primary training table.
- `sentences`: derived sentence spans; redundant for current token-window training.
- `legacy.news_agency_as_source`: thesis-era document-level source-attribution provenance. It mixes QIDs with sentinels such as `_`, `unk`, and `NIL`, and is not part of the current annotation target.
- `entities[].entity_id`, `entities[].nel`, `entities[].normalized_surface`, `entities[].has_ocr_correction`, `entities[].max_ocr_levenshtein`, `entities[].label_original`, and `entities[].status`: legacy normalization/linking/OCR/audit fields. The current label is `entities[].label`; current entity links use `entities[].wikidata_url`; compact OCR corrections use `entities[].ocr_correction`.

Workbench/audit artifacts such as curation operation files, TSV materializations, tokenization migration reports, and model-specific quality diagnostics are intentionally not part of the Hugging Face dataset payload.

## Loading

```python
from datasets import load_dataset

dataset = load_dataset(
    "impresso-project/impresso-mediaagencies-ner-dataset",
    data_files={
        "train": "data/train.jsonl",
        "validation": "data/validation.jsonl",
        "test": "data/test.jsonl",
    },
)
```

## Label Policy

Trainable labels use two namespaces:

- `org.ent.pressagency.<canonical_id>`
- `org.ent.radiostation.<canonical_id>`

Forbidden legacy labels such as `unk`, unresolved bare `ag`, `org.ent.pressagency.ag`, and `pers.ind.articleauthor` are excluded from the trainable label space.

The source copy of this card lives in the workbench. Publishing scripts copy it into the Hugging Face dataset repository.

## License

License metadata is set to `other` until the final redistribution terms for the converted annotations and Impresso text snippets are confirmed.