Tibetan Metadata RoBERTa NER
Fine-tuned spsither/tibetan_RoBERTa_S_e3 for title and author span detection in Tibetan text segments.
Labels (BIO)
| ID | Label |
|---|---|
| 0 | O |
| 1 | B-TITLE |
| 2 | I-TITLE |
| 3 | B-AUTHOR |
| 4 | I-AUTHOR |
Training
- Dataset: ganga4364/tibetan-metadata-detector
- 3 epochs, batch size 64, lr 2e-5
- Sliding-window examples: 512 tokens, stride 256, 15B+15E overlap-aware
Test metrics (full test split, 108,954 windows)
| Metric | Value |
|---|---|
| span F1 | 14.9% |
| span precision | 11.9% |
| span recall | 20.0% |
| title F1 | 16.5% |
| author F1 | 1.4% |
Inference
Important: Do not pass entire long segments in one forward pass. Use the same sliding-window pipeline as training (15 begin + 15 end, stride 256), merge overlapping predictions, then extract spans.
from pipeline.inference import load_model_and_tokenizer, predict_segment, highlight_spans
model, tokenizer, device = load_model_and_tokenizer("ganga4364/tibetan-metadata-roberta-ner")
spans = predict_segment(model, tokenizer, your_tibetan_text, device=device)
print(highlight_spans(your_tibetan_text, spans))
Demo
Try the Gradio Space: ganga4364/tibetan-metadata-highlight