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 11 columns:

column	type	description
source	string	source-language sentence
target	string	target-language sentence
source_lang	string	ISO-639-3 source language code (from upstream metadata)
target_lang	string	ISO-639-3 target language code (from upstream metadata)
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
source_lang_claude_haiku_detect	string	ISO-639-3 of language Haiku detected in source (or unk)
source_lang_claude_haiku_detect_confidence	float32	Haiku's self-reported confidence in [0.0, 1.0] for the source-lang detection
target_lang_claude_haiku_detect	string	ISO-639-3 of language Haiku detected in target (or unk)
target_lang_claude_haiku_detect_confidence	float32	Haiku's self-reported confidence in [0.0, 1.0] for the target-lang detection

Keep / reject criteria

A row is kept (in natgillin/translations) when all hold:

1. claude_haiku_score > 0.8
2. NOT (source_lang_claude_haiku_detect != source_lang AND source_lang_claude_haiku_detect_confidence >= 0.70)
3. NOT (target_lang_claude_haiku_detect != target_lang AND target_lang_claude_haiku_detect_confidence >= 0.70)

I.e. a language mismatch only rejects a row when Haiku is confident about the mismatch (>=0.70). Low-confidence detections (typical for low-resource languages like Aymara, Quechua, Guarani) do NOT auto-reject, since Haiku's training has thin coverage there. Both labels and their confidences are stored on every row so downstream consumers can re-derive the split.

How xxhash_intdigest is computed

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 + lang-detect columns are computed

Rows are batched and sent to Claude Haiku 4.5 with the prompt below. For each pair, the model returns three values:

• s: quality score in [0.0, 1.0] → stored as claude_haiku_score
• sl: detected source language (ISO-639-3) → stored as source_lang_claude_haiku_detect
• tl: detected target language (ISO-639-3) → stored as target_lang_claude_haiku_detect

Exact prompt template

You are a translation judge. For each pair, return FIVE values:
- s: fluency+faithfulness score 0.00-1.00 (1=perfect, 0.8=cutoff, 0.6=awkward, 0.4=broken, 0.2=garbled, 0=empty/wrong)
- sl: detected language of SRC text (ISO-639-3, 3 lowercase letters, or 'unk' if unsure)
- slp: confidence 0.00-1.00 that `sl` is correct (low if you're guessing or the language is low-resource like aym/que/grn)
- tl: detected language of TGT text (ISO-639-3, 3 lowercase letters, or 'unk')
- tlp: confidence 0.00-1.00 for `tl`

Metadata claims: source={src_lang}, target={tgt_lang}. Use the actual language you observe, not the metadata. Report your true confidence — use slp/tlp < 0.5 when the language is hard to identify rather than guessing.

Pairs:
{pairs_block}

Return ONLY this JSON (no prose), exactly {n} items in order:
{{"results":[{{"s":0.95,"sl":"aym","slp":0.4,"tl":"eng","tlp":0.99}}, ...]}}

JSON schema sent alongside the prompt

{
  "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.