gliner_lviv

This model is a fine-tuned version of [medieval-data/gliner_multi-v2.1-medieval-latin] on historical Latin records from Lviv, given by the Center for Digital Humanities.

F1 Score 0.8500 Precision 0.8547 Recall 0.8453

It is optimized for extracting person names (person) from complex historical Latin syntax using span-level entity representation.

Model description

gliner_lviv is a Generalist and Lightweight Named Entity Recognition (GLiNER) model adapted for historical Latin archival documents. Unlike traditional sequence-tagging models (like BERT/RoBERTa using token-level token classification), GLiNER formulates NER as a bidirectional span-matching task between input text sequences and target label representations.

Key features:

• Architecture: Bi-encoder transformer backbone fine-tuned for span-level entity boundary detection.
• Task: Span-based Named Entity Recognition (Strict Boundary Extraction).
• Target Entities: person (Latin historical names, patronymics, and compound titles).
• Domain: Medieval and Early Modern Latin legal and administrative council documents.

Intended uses & limitations

Intended uses

• Automated extraction and prosopographical indexing of historical figures, witnesses, notaries, and council members in Latin archival acts.
• Assisting digital humanities researchers in indexing unstandardized early modern municipal manuscripts.
• Zero-shot / Few-shot transfer on related medieval and early modern Latin administrative corpora.

Limitations

• Context length: Optimized for sliding-window chunks up to 250–384 tokens; document-level inputs should be segmented before inference.
• Scribal abbreviations: Unexpanded scribal abbreviations (e.g., = ligatures, tildes, suspended patronymics) may require normalized transcriptions for highest recall.
• Domain specificity: Primarily trained on municipal administrative and legal formulations; performance may vary on liturgical or literary Latin texts.

Training and evaluation data

• Training Data (data_ai): Transcriptions of historical acts and books (dataset_1_book, dataset_23_book) containing inline ground-truth <per>...</per> person entities. Segmented into period-bounded sentence units and converted into token-span indices.
  • Split: 80% training, 20% validation.
• Hold-Out Benchmark (test_data): Unseen, isolated historical document transcriptions reserved strictly for final entity-level F1 benchmarking.

Training procedure

Training hyperparameters

The following hyperparameters were used during fine-tuning:

• learning_rate: 3e-05 (with linear warmup and decay)
• train_batch_size: 4
• eval_batch_size: 8
• gradient_accumulation_steps: 2 (effective batch size = 8)
• warmup_ratio: 0.1
• max_grad_norm: 1.0
• weight_decay: 0.01
• num_epochs: 3
• metric_for_best_model: eval_loss
• load_best_model_at_end: True

Training results

Epoch	Step	Training Loss	Validation Loss
1.0	67	No log	16.1674
2.0	134	No log	14.7571
3.0	201	No log	15.5742

---