Datasets:
Fix ara-deu qrels using original MLQA contexts
This pull request was prepared and validated by OpenAI Codex at the reporter's request.
Related report: https://github.com/embeddings-benchmark/mteb/issues/5585
The ara-deu subset currently links almost every German query to an unrelated Arabic document. This PR recovers its complete relevance sets from the original MLQA release and replaces exactly two files:
ara-deu-qrels/test-00000-of-00001.parquetara-deu-qrels/validation-00000-of-00001.parquet
Query/corpus texts and IDs, other language pairs, Arrow column types, Hugging Face feature metadata, and positive score values remain unchanged. The new files retain Snappy compression and the existing row-group sizes.
Recovery and validation
Base revision: cf59ddd8f4aaf39ce1869361e09252698c340945.
Original release: https://dl.fbaipublicfiles.com/MLQA/MLQA_V1.zip
ZIP SHA256: 246e8089933d13007fe80684d5c5c0713d6834cf8b3b4a0ec7c66f0a0d2baac8.
For both splits, exact German question strings were mapped to all corresponding original Arabic context strings, then to the existing corpus IDs. Duplicate question text is handled by retaining every original positive context.
| Split | Queries | Positive rows | Incorrect rows before | Incorrect rows after |
|---|---|---|---|---|
| test | 1,648 | 1,649 | 1,648 | 0 |
| validation | 207 | 207 | 207 | 0 |
Validation compares complete per-query relevance sets, not just whether each positive row is valid. Every query now has exactly its recovered relevance set; there are no missing source questions/contexts, invalid IDs, or duplicate edges. Both Parquet files were read back and checked against the original source, with schema and metadata equality verified. Test query Q932 correctly retains two positives, C1353 and C436; the misplaced extra positive for Q353 is removed. All other 48 pairs passed the original-source audit in both splits before this two-file patch.
Independent reproduction
The 32-line public script in issue #5585 can be run with REV set to this PR's commit SHA (shown under Files and commits). It downloads the source ZIP and the six ara-deu files. Expected output after the repair:
test queries 1648 qrel_rows 1649 wrong 0
validation queries 207 qrel_rows 207 wrong 0
The same script against the base revision reports 1,648 and 207 incorrect rows. No model or GPU is needed.
For impact, CPU rescoring of unchanged cached intfloat/multilingual-e5-large embeddings/rankings changes test nDCG@10 from 0.0032385779 to 0.6951879989 when only the relevance mappings are repaired (published historical score: 0.69496). This was not a fresh model inference run.
After merging this dataset repair, MTEB's task should pin the resulting corrected dataset revision, so affected and repaired benchmark results remain distinguishable.
responded in github issue