---
language:
- multilingual
tags:
- translation
- parallel-text
- mtdata
- opus
license: mit
---

# natgillin/translations-raw

**Frozen, canonical raw bitext** consolidated from upstream `alvations/mtdata-raw*` snapshots (since deleted). This is the read-only source-of-truth for downstream quality-filtering pipelines.

- **31,663 parquet files** (1566.8 GB)
- **49 language pairs** under `data/<src-tgt>/`
- Schema: 5 columns — see below
- **Read-only** for downstream pipelines. Do not delete or modify.

## Schema

Each parquet has 5 columns:

| column | type | description |
|---|---|---|
| `source` | string | source-language sentence |
| `target` | string | target-language sentence |
| `source_lang` | string | ISO-639-3 source language code |
| `target_lang` | string | ISO-639-3 target language code |
| `origin` | string | upstream OPUS / mtdata corpus tag (e.g. `opus-nllb`, `opus-paracrawl`) |

## Layout

```
data/
  aym-eng/
    opus-nllb-mtdata-recovery-00000-of-00200.parquet
    opus-globalvoices-v2018q4-mtdata-recovery-00000-of-00200.parquet
    ...
  aze-eng/
    opus-nllb-mtdata-recovery-00000-of-00200.parquet
    ...
  bel-eng/
  bul-eng/
  ces-eng/
  ...
```

49 pairs total. File naming convention:
- `<origin>-NNNNN-of-NNNNN.parquet` (original consolidation runs)
- `<origin>-mtdata-recovery-NNNNN-of-NNNNN.parquet` (re-sourced from mtdata)

## Provenance

- Originally consolidated from `alvations/mtdata-raw`, `alvations/mtdata-raw-2`, `alvations/mtdata-raw-hashed` (now deleted).
- Files marked `-mtdata-recovery-` were re-pulled directly from the [mtdata](https://github.com/thammegowda/mtdata) library on 2026-06-09.

## Downstream filtered versions

- [`natgillin/translations`](https://huggingface.co/datasets/natgillin/translations) — rows that scored `> 0.8` on Claude Haiku 4.5 translation-quality evaluation
- [`natgillin/translations-rejected`](https://huggingface.co/datasets/natgillin/translations-rejected) — rows that scored `<= 0.8`