---
language:
- ro
license: mit
task_categories:
- text-generation
tags:
- diacritics
- romanian
- adr
- text-restoration
size_categories:
- 100K<n<1M
---

# Romanian Diacritic Restoration Dataset (dexonline)

Data splits for Romanian automatic diacritic restoration (ADR), derived from the dexonline lexicographic corpus.

## Dataset Description

Each example contains a `input` (stripped text without diacritics), a `target` (gold-standard diacritized text), and an `id`.

- **Source**: [dexonline.ro](https://dexonline.ro) -- open-access Romanian lexicographic resource
- **Training pairs**: 299,633 examples (NFKD diacritic stripping + cedilla normalization)
- **Validation**: 33,292 examples
- **Test sets**: CRAWLER-1000 and DLRLC-1000 (clean + 3 noise levels each = 8 test files)

## Splits

| Split | Items | Description |
|-------|-------|-------------|
| `train.jsonl` | 299,633 | Training pairs (deduplicated against test sets) |
| `val.jsonl` | 33,292 | Validation pairs |
| `test_crawler_1000_clean.jsonl` | 1,000 | Clean web-crawled literary text |
| `test_crawler_1000_{low,medium,high}.jsonl` | 1,000 each | Noisy variants (typos, casing, diacritic corruption) |
| `test_dlrlc_1000_clean.jsonl` | 1,000 | Clean lexicographic definitions |
| `test_dlrlc_1000_{low,medium,high}.jsonl` | 1,000 each | Noisy variants |

## Noise Levels

- **Low**: 1% typos, 0.5% case flips, 5% diacritic dropping, 10% cedilla mixing
- **Medium**: 3% typos, 1% case flips, 15% diacritic dropping, 2% insertion, 30% cedilla mixing
- **High**: 5% typos, 2% case flips, 30% diacritic dropping, 5% insertion, 50% cedilla mixing

## Usage

```python
from datasets import load_dataset
ds = load_dataset("klusai/diacritics-ro", data_files={"train": "data/train.jsonl", "test": "data/test_crawler_1000_clean.jsonl"})
```

## Associated Paper

"Comparing Five Model Families for Romanian Diacritic Restoration: From Dictionary Lookup to LoRA-Adapted Decoders" (Nadas, 2026)

## Code

[github.com/klusai/diacritics-finetuning-code](https://github.com/klusai/diacritics-finetuning-code)