| --- |
| language: |
| - multilingual |
| tags: |
| - translation |
| - quality-estimation |
| - claude-haiku |
| license: mit |
| --- |
| |
| # natgillin/translations — Claude-Haiku-filtered bitext |
|
|
| Rows from `natgillin/translations-raw` that scored `> 0.8` on Claude Haiku 4.5 translation-quality evaluation. Globally deduplicated by xxh3-64 of `source\ntarget`. |
|
|
| ## Schema |
|
|
| Each parquet file has 7 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 corpus tag (e.g. `opus-nllb`) | |
| | `xxhash_intdigest` | uint64 | hash of the pair (see below) | |
| | `claude_haiku_score` | float32 | quality score in `[0.0, 1.0]` from Claude Haiku 4.5 | |
|
|
| The split is `claude_haiku_score > 0.8`: kept rows live in `natgillin/translations`, rejected rows in `natgillin/translations-rejected`. |
|
|
| ## How `xxhash_intdigest` is computed |
| |
| ```python |
| import xxhash |
| |
| def row_hash(source: str, target: str) -> int: |
| return xxhash.xxh3_64(f"{source}\n{target}".encode("utf-8")).intdigest() |
| ``` |
| |
| The hash is the **xxh3-64** intdigest of `f"{source}\n{target}"` encoded as UTF-8. |
| It is stable across runs and lets you deduplicate or join rows by content. |
|
|
| ## How `claude_haiku_score` is computed |
|
|
| Rows are batched (default 20 / call) and sent to Claude Haiku 4.5 with the prompt below. |
| The model returns a JSON array of floats in `[0.0, 1.0]`; we attach each float to its |
| corresponding row as `claude_haiku_score`. |
|
|
| ### Exact prompt template |
|
|
| ``` |
| You are a translation quality judge. Rate each TARGET translation's fluency and faithfulness to its SOURCE on a scale 0.00 to 1.00. |
| |
| Rubric: |
| - 1.00: native-fluent target, accurately conveys source meaning, no errors |
| - 0.80: mostly fluent, minor errors that don't impede understanding (CUTOFF) |
| - 0.60: understandable but awkward; some meaning loss |
| - 0.40: broken grammar or significant meaning drift |
| - 0.20: mostly garbled / barely comprehensible |
| - 0.00: empty, wrong language, or not a translation at all |
| |
| Source language: {src_lang}. Target language: {tgt_lang}. |
| |
| Translations (return scores in the same order): |
| {pairs_block} |
| |
| Output ONLY a JSON array of exactly {n} floats between 0.0 and 1.0. No prose. Example: [0.95, 0.4, 0.85, ...] |
| ``` |
|
|
| ### JSON schema sent alongside the prompt |
|
|
| ```json |
| { |
| "type": "object", |
| "properties": { |
| "scores": { |
| "type": "array", |
| "items": { |
| "type": "number", |
| "minimum": 0.0, |
| "maximum": 1.0 |
| } |
| } |
| }, |
| "required": [ |
| "scores" |
| ], |
| "additionalProperties": false |
| } |
| ``` |
|
|
| ## Source |
|
|
| All rows come from `natgillin/translations-raw` (an OPUS / mtdata mirror). |
| Before any Haiku call, rows are deduplicated globally by `xxhash_intdigest` |
| so identical pairs are scored exactly once. |
|
|
