mmBERT Impresso Media Sources NER
Fine-tuned token-classification model for media-source mention recognition in historical Impresso newspaper text.
The v2.0.1 model predicts cited source mentions for:
- news agencies, labelled as
org.ent.pressagency.<canonical_id> - radio stations, labelled as
org.ent.radiostation.<canonical_id>
The repository name uses "media sources" because the label space may later grow to cover other cited media-source families.
Provenance
- Dataset:
impresso-project/impresso-mediaagencies-ner-dataset - Dataset revision:
v2.0.0 - Dataset commit:
a7ac5dc1ec0dd92ae848dbccd258aa0361830da3 - Training code commit: not recorded by this completed run
- Publication/workbench commit:
429a3767288abd16a43b2f7996453f73df350c7c - Model revision:
v2.0.1 - Model commit:
5f600305a86b90d5467c5d71b827e0f33872776c - Previous model commit:
9899ad960b9bc310ee51cd7ee658fd3882b6b140 - Base model:
impresso-project/mmbert-multilingual-impresso-continued-mlm - Upstream base:
jhu-clsp/mmBERT-base - License: MIT
The model was trained from the pinned published Hugging Face dataset commit above. The selected checkpoint was chosen on validation performance, not on the held-out test set. The completed training run did not record a Git SHA in training_start_report.json; the publication/workbench commit is therefore recorded separately and should not be treated as the exact training-code commit. The v2.0.1 revision updates the Hugging Face custom pipeline runtime for the existing v2.0.0 trained weights; no retraining was performed.
Training Setup
- Annotation tokenization:
unicode-word-punctuation-v1 - Subtoken supervision: all subtokens labelled, with B-to-I conversion for continuation subtokens
- Decoder used for evaluation/deployment:
first_subtoken_viterbi - Adaptation: top 4 ModernBERT layers unfrozen
The model config.json records annotation_tokenization, label_all_tokens, subtoken_labeling, and subtoken_decoding. Inference must apply the declared annotation-token profile and the declared subtoken_decoding policy. label_all_tokens describes training supervision; it does not by itself define inference decoding.
Evaluation
Exact entity metrics use the constrained first_subtoken_viterbi decoder over the checkpoint label space.
Validation metrics for the selected checkpoint:
- Precision: 0.9398
- Recall: 0.9174
- F1: 0.9285
- Correct/gold/predicted entities: 500 / 545 / 519
Held-out test metrics:
- Precision: 0.9346
- Recall: 0.8644
- F1: 0.8981
- Correct/gold/predicted entities: 529 / 612 / 566
Runtime inference validation:
- Release-artifact inference was verified on the complete v2.0.0 held-out test set for the published v2.0.1 runtime artifact.
- The Hugging Face runtime pipeline matched the evaluator's decoded BIO sequence for 458 / 458 documents.
- Recomputed runtime metrics matched the evaluator exactly, including 529 / 612 / 566 exact entities and F1 0.8981324278.
Intended Use
Use this model to identify cited news-agency and radio-station source mentions in historical newspaper text. The model is intended for corpus enrichment and assisted curation workflows, not as a general-purpose named-entity recognizer.
The input text should be tokenized with the same annotation-tokenization policy used by the training data. For deployment through the custom pipeline files in this repository, use the model's configured first_subtoken_viterbi decoding policy.