DanielDDDS's picture
Clean upload: dictabert-large P1, relaxed F1=62.6%
c84096a verified
---
language: he
tags:
- token-classification
- hebrew
- recipe
- ner
license: mit
---
# Hebrew Recipe Modification NER
DictaBERT-large fine-tuned for recipe modification extraction from Hebrew YouTube comments.
Trained with class weighting (P1) on silver labels from a 3-pass LLM teacher pipeline (v2).
## Labels
- `B/I-SUBSTITUTION` โ€” ingredient substitution
- `B/I-ADDITION` โ€” ingredient addition
- `B/I-QUANTITY` โ€” quantity change
- `B/I-TECHNIQUE` โ€” technique change
## Usage
```python
from transformers import pipeline
pipe = pipeline("token-classification",
model="DanielDDDS/hebrew-recipe-modification-ner",
aggregation_strategy="simple")
pipe("ืืคืฉืจ ืœื”ื—ืœื™ืฃ ื—ืžืื” ื‘ืฉืžืŸ ืงื•ืงื•ืก")
```
## Performance (corrected gold test set, n=496, 38 spans)
- Exact Entity F1: 25.5%
- Relaxed Entity F1: 62.6%
- Model: DictaBERT-large + linear head, class weights (P1)
- Beats LLM teacher on relaxed F1 (teacher: 48.4%)