---
language: [ro]
license: mit
tags: [diacritics, romanian, bilstm]
pipeline_tag: text-generation
---
# Character BiLSTM for Romanian Diacritic Restoration

2-layer bidirectional LSTM (2.4M parameters) for character-level diacritic restoration.

## Results (CRAWLER-1000 clean)
| Metric | Value |
|--------|-------|
| Word Accuracy | 96.23% |
| DER | 0.033 |
| Speed | 111 sent/s |

**Noise robustness**: Collapses under noise (96% to 37% at high noise). Best for clean text only.

## Training
- Data: 299,633 examples from dexonline corpus
- Hardware: Apple M3 Ultra (MPS)
- Epochs: 5, batch size 256, Adam lr=1e-3
- Training time: ~45 minutes

## Resources
- Paper: "Comparing Five Model Families for Romanian Diacritic Restoration"
- Dataset: [klusai/diacritics-ro](https://huggingface.co/datasets/klusai/diacritics-ro)
- Code: [github.com/klusai/diacritics-finetuning-code](https://github.com/klusai/diacritics-finetuning-code)