impresso-project/mmbert-impresso-mediasources-ner

mmBERT Impresso Media Sources NER

Source model card for the future Hugging Face model repository impresso-project/mmbert-impresso-mediasources-ner.

The published model should include:

• model weights
• tokenizer files
• config.json
• requirements.txt
• this model card

The first publication is a standard Transformers token-classification model. A custom Impresso JSON pipeline will be added later, after the inference interface has been finalized.

The v0.1 model predicts cited source mentions for news agencies and radio stations. The repository name uses "media sources" because the label space may later grow to cover other cited media-source families, such as newspaper citations.

This model is fine-tuned from impresso-project/mmbert-multilingual-impresso-continued-mlm, which continued pretraining from jhu-clsp/mmBERT-base. The upstream mmBERT base model is released under the MIT license.

Evaluation

Validation/dev exact entity metrics:

• Precision: 0.8561
• Recall: 0.9040
• F1: 0.8794
• Correct/gold/predicted entities: 113 / 125 / 132

Test exact entity metrics:

• Precision: 0.9111
• Recall: 0.9213
• F1: 0.9162
• Correct/gold/predicted entities: 164 / 178 / 180

These scores are from the legacy converted French/German HIPE-style evaluation splits. Manual disagreement curation is planned after the first publication.