Datasets:
Remove FinEntity test-set documents (benchmark contamination)
This PR removes the FinEntity documents that are used as test items by the Decision Index benchmark suite, and adds a short note to the README.
Why: benchmark contamination. The suite's FinEntity test documents (979 items, 968 distinct documents, which is the full public FinEntity release) appear in this dataset's training and validation splits, so models trained on it get inflated FinEntity scores. The README note asks that future training on this dataset exclude these documents.
Method: whitespace- and case-normalised exact match of each row's state against the documents of the Decision Index frozen suite. long_context and views_pt_en were also checked with substring matching and through the id link back to their source core row; none of their rows derive from FinEntity. eval/* was checked the same way.
Rows removed:
| File | Before | Removed | After |
|---|---|---|---|
data/core/train.jsonl |
21,150 | 1,911 | 19,239 |
data/core/validation.jsonl |
1,221 | 83 | 1,138 |
data/core/heldout.jsonl |
1,191 | 0 | 1,191 |
data/long_context/train.jsonl |
1,517 | 0 | 1,517 |
data/long_context/validation.jsonl |
59 | 0 | 59 |
data/views_pt_en/train.jsonl |
1,619 | 0 | 1,619 |
eval/* |
2,268 | 0 | 2,268 |
All 1,994 FinEntity-derived rows matched a suite document, so no FinEntity rows remain. Kept rows are byte-identical and in their original order. MANIFEST.sha256 is updated for the changed files.
Based on revision 26d9a680.