---
annotations_creators:
- found
language_creators:
- found
language:
- ar
- de
- en
- es
- fr
- hi
- id
- it
- ja
- ko
- nl
- pl
- pt
- ru
- sv
- th
- tr
- uk
- vi
- zh
license:
- apache-2.0
multilinguality:
- multilingual
pretty_name: UTS-Datasets
size_categories:
- 10M<n<100M
source_datasets:
- opus-100
- opus_ccmatrix
- opus_paracrawl
- open_subtitles
- wmt19
- wmt20
- wmt21
task_categories:
- translation
task_ids:
- translation
---

# UTS-Datasets

Processed parallel training and validation data for the [Universal Translation System](https://github.com/code-with-zeeshan/universal-translation-system). Generated by the UTS data pipeline with quality enhancement stages.

## Dataset Contents

### Files

| File | Description |
|------|-------------|
| `train_final.txt` | ~4M parallel sentences (all language pairs, sampled & augmented) |
| `val_final.txt` | ~100K validation sentences (held out from training) |
| `vocab/` | Script-grouped SentencePiece vocabulary packs (6 groups × 32K tokens) |
| `pipeline_state.json` | Pipeline checkpoint state for resumability |

### Language Coverage

20 languages, 190 directional pairs:

| Group | Languages |
|-------|-----------|
| Latin | en, es, fr, de, it, pt, nl, sv, pl, id, vi, tr |
| CJK | zh, ja, ko |
| Arabic | ar |
| Devanagari | hi |
| Cyrillic | ru, uk |
| Thai | th |

### Training Distribution (top pairs)

| Pair | Sentences |
|------|-----------|
| en-es | 400K |
| en-fr | 400K |
| en-de | 400K |
| en-zh | 300K |
| en-ru | 300K |
| en-ja | 200K |
| en-ar | 200K |
| en-pt | 200K |
| en-it | 200K |
| en-hi | 100K |
| Others (10 pairs) | 60K-100K each |
| Non-English pivots (14 pairs) | 40K each |

## Data Processing Pipeline

The data was processed through the UTS pipeline stages:

1. **Download** — OPUS-100 + supplementary sources (CCMatrix, ParaCrawl, OpenSubtitles, WMT)
2. **Sampling** — Length filtering (16-128 tokens), language detection, deduplication
3. **Augmentation** — False friends replacement, idiom insertion, pivot backtranslation
4. **Knowledge Distillation** — NLLB-3.3B teacher → soft labels (GPU-only stage)
5. **Quality Filtering** — COMET-22 neural quality scoring (threshold 0.7)
6. **Vocabulary** — SentencePiece training per script group (unigram model, 32K tokens)

## Usage

```bash
# Download via HF Hub
pip install huggingface_hub
huggingface-cli download code-with-zeeshan/UTS-Datasets --repo-type dataset --local-dir ./uts_data

# Use with UTS training
uts train --full --config config/override/my_config.yaml --hub-repo-id code-with-zeeshan/UTS-Datasets
```

## Data Format

Each line in `train_final.txt` / `val_final.txt`:
```
<langpair> ||| <source_text> ||| <target_text>
```

Example:
```
en-es ||| Hello world ||| Hola mundo
```

## Recreating This Dataset

```bash
uts config --interactive
# Enable all stages: download, sample, augment, create_ready,
#   validate, vocab, wiki_bt, knowledge_distillation, comet
uts data --pipeline --config config/override/my_config.yaml --hub-repo-id code-with-zeeshan/UTS-Datasets
```

## License

Apache 2.0