bowang0911/finqa-hard-negatives

FinQA hard negatives

Near-duplicate hard negatives for training financial-document retrievers, built from the train split of FinQA (Chen et al., arXiv:2109.00122).

Each row is one question, its gold evidence unit(s), and the other evidence units from the same filing page as hard negatives — same company, same fiscal year, same boilerplate. The distinction a model must learn is which sentence or table row on the page answers the question, not which company or topic is relevant.

Files

file	rows	negatives/query	notes
finqa_hn_full.jsonl	6,239	24.96 (3–91)	all page-mates
finqa_hn8.jsonl	6,239	7.89 (3–8)	hardest 8 only

Schema (one JSON object per line):

{
  "query": "what was the percentage change in net cash from operating activities from 2008 to 2009?",
  "document": ["gold evidence unit", "..."],
  "hard_negatives": ["page-mate 1", "page-mate 2", "..."]
}

document is a list: 1.71 golds per query on average, multi-positive preserved rather than exploded into duplicate rows.

Negatives are ordered hardest-first, using FinQA's own text_retrieved_all / table_retrieved_all retriever scores. finqa_hn8.jsonl is the first 8 of that ordering. Which file you want depends on your sampler: if it draws negatives uniformly at random rather than top-k, the truncated file is what actually controls hardness.

Construction

Both variants derive from FinQA dataset/train.json (6,251 records). Filters applied, in order:

1. One renderer for both sides. FinQA's table_row_to_text does not reproduce its own tokenizer spacing on 263 table golds. Positives and negatives go through a single renderer — mixing FinQA's string for positives with ours for negatives would leave a formatting shortcut that a model can exploit instead of learning the content.
2. Negative-index gold fixed. text_-1 (RE/2010/page_120) is a Python negative index resolving to all_text[23]; unhandled, that gold lands in its own distractor list.
3. Dropped: 9,567 OCR filler distractors ("."), 71 verbatim copies of the row's own gold, 15,300 sibling/cross-page golds.
4. Sibling-gold filter at two scopes — table/ind matches page-scoped, text matches corpus-wide, because boilerplate like "u.s ." recurs verbatim across filings. This took gold-in-pooled-hard-negatives from 23% of batches to 0%.
5. --min-negs 3: dropped 12 rows that had only 2 negatives (0.19%).

Known limitations

• Test-company overlap. 99 of the 100 companies in FinQA's test split also appear in train. Grade gains on FinanceBench or TAT-QA, not on FinQA itself.
• Repeated positives. ~14% of batches still contain the same gold twice (two questions citing one sentence). Fixable by merging into multi-positive rows, but that changes the query distribution.
• Some golds are junk. A few FinQA gold units are fragments ("u.s .", "12 .").
• Below common dataset-size floors. 6,239 rows; some training mixes expect ≥8,192 per dataset, in which case merge with another finance source.
• Trainers that pool hard negatives batch-wide make one row's negatives everyone's negatives. The sibling-gold filter above is what makes that safe here; an exact-string dedup at collate time is still worth adding.

License

Derived from FinQA; consult the upstream repository for its terms. FinQA is itself built on earnings reports from public SEC filings.

Citation

@article{chen2021finqa,
  title={FinQA: A Dataset of Numerical Reasoning over Financial Data},
  author={Chen, Zhiyu and Chen, Wenhu and Smiley, Charese and Shah, Sameena and
          Borova, Iana and Langdon, Dylan and Moussa, Reema and Beane, Matt and
          Huang, Ting-Hao and Routledge, Bryan and Wang, William Yang},
  journal={arXiv preprint arXiv:2109.00122},
  year={2021}
}