---
language:
  - ro
license: cc-by-sa-4.0
task_categories:
  - fill-mask
tags:
  - romanian
  - educational
  - fill-in-the-blank
  - wikipedia
  - academic
size_categories:
  - 10K<n<100K
---

# EduRO-FillBlank

A Romanian-language dataset of **72,541 fill-in-the-blank examples** 
extracted from Romanian Wikipedia, covering 45 academic domains.
Built to support fine-tuning of masked language models for 
automatic educational quiz generation.

## Dataset Summary

| Property | Value |
|---|---|
| Language | Romanian |
| Domain | Academic (university-level) |
| Total examples (raw) | 100,212 |
| Total examples (cleaned) | 72,541 |
| Sources | 4,364 Romanian Wikipedia articles |
| Categories | 45 academic domains |
| License | CC BY-SA 4.0 |

## Splits

| Split | File | Examples |
|---|---|---|
| Train | eduro_v2_train.csv | 58,032 |
| Validation | eduro_v2_val.csv | 7,254 |
| Test | eduro_v2_test.csv | 7,255 |

## Dataset Structure

| Column | Type | Description |
|---|---|---|
| `sentence` | string | Original sentence from Wikipedia |
| `masked_sentence` | string | Sentence with target word replaced by `_____` |
| `answer` | string | The masked word (correct answer) |
| `category` | string | Academic domain (e.g., Biologie, Informatică) |
| `source` | string | Wikipedia article title |

## Example

```json
{
  "sentence": "Fotosinteza este procesul biochimic prin care plantele produc glucoza.",
  "masked_sentence": "Fotosinteza este procesul _____ prin care plantele produc glucoza.",
  "answer": "biochimic",
  "category": "Biologie",
  "source": "Fotosinteza"
}
```

## How to Load

```python
from datasets import load_dataset

dataset = load_dataset("dariurs/EduRO-FillBlank")
train = dataset["train"]
```

## Construction Pipeline

1. **Collection** — Romanian Wikipedia API, 45 academic categories, 4,364 articles
2. **Masking** — scoring algorithm prioritizing content words
3. **Cleaning** — 6-stage pipeline:
   - Filter 1: Sub-token removal (−5,112)
   - Filter 2: Frequency filtering (−16,631)
   - Filter 3: Stopword removal (−2,463)
   - Filter 4: Sentence quality (−3,386)
   - Filter 5: Answer consistency (−68)
   - Filter 6: Deduplication (−11)
4. **Split** — 80/10/10 train/val/test

## Associated Model

`dariurs/eduro-bert-romanian-fillblank-v3`

## Citation

```
[to be added after paper publication]
```