README.md

metadata

license: mit
language:
  - en
task_categories:
  - question-answering
  - table-question-answering
tags:
  - financial
  - numerical-reasoning
  - table-qa
  - earnings-reports
size_categories:
  - 1K<n<10K
dataset_info:
  features:
    - name: id
      dtype: string
    - name: filename
      dtype: string
    - name: pre_text
      list: string
    - name: post_text
      list: string
    - name: table
      list:
        list: string
    - name: table_ori
      list:
        list: string
    - name: question
      dtype: string
    - name: answer
      dtype: string
    - name: exe_ans
      dtype: string
    - name: explanation
      dtype: string
    - name: ann_table_rows
      list: int32
    - name: ann_text_rows
      list: int32
    - name: steps
      struct:
        - name: op
          list: string
        - name: arg1
          list: string
        - name: arg2
          list: string
        - name: res
          list: string
    - name: program
      dtype: string
    - name: program_re
      dtype: string
    - name: gold_inds
      dtype: string
  splits:
    - name: train
      num_bytes: 32780740
      num_examples: 6251
    - name: validation
      num_bytes: 4573763
      num_examples: 883
    - name: test
      num_bytes: 5911480
      num_examples: 1147
  download_size: 39700298
  dataset_size: 43265983
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*

FinQA

A full-fidelity repackaging of the FinQA dataset (Chen et al., EMNLP 2021) for numerical reasoning over financial tables.

FinQA contains questions over earnings reports from S&P 500 companies (1999–2019), sourced from the FinTabNet dataset. Each example pairs a financial table and surrounding text with a question, a human-readable answer, and a structured reasoning program that specifies the arithmetic operations needed to derive the answer.

Why this version

The existing HuggingFace copy (dreamerdeo/finqa) omits program, program_re, and steps — the reasoning annotations that are FinQA's primary contribution. This version preserves the complete schema from the canonical GitHub release, excluding only model-artifact fields (table_retrieved*, text_retrieved*, qa.tfidftopn, qa.model_input) that were bundled with the original retriever checkpoint.

Splits

Split	Examples
train	6,251
validation	883
test	1,147

The private leaderboard test set (no ground-truth references) is intentionally excluded.

Schema

Field	Type	Description
id	string	Unique example ID: {TICKER}/{YEAR}/page_{N}.pdf-{idx}
filename	string	Source PDF path: {TICKER}/{YEAR}/page_{N}.pdf
pre_text	list[string]	Text passages before the table
post_text	list[string]	Text passages after the table
table	list[list[string]]	Normalized table (numbers unformatted, HTML stripped)
table_ori	list[list[string]]	Original table with HTML formatting (superscripts, comma-separated numbers)
question	string	The financial question
answer	string	Human-readable answer (e.g. "93.5%")
exe_ans	string	Raw execution result as string; numeric (e.g. "0.935") or "yes"/"no"
explanation	string	Free-text explanation; sparse (~16% non-empty)
ann_table_rows	list[int]	Zero-indexed table rows annotated as gold evidence
ann_text_rows	list[int]	Zero-indexed text passages annotated as gold evidence
steps	struct	Structured execution steps (see below)
program	string	Flat DSL program with #N back-references (e.g. subtract(920, 95), divide(#0, 5))
program_re	string	Fully nested program form (e.g. divide(subtract(920, 95), 5))
gold_inds	string	JSON-serialized dict mapping evidence keys to text (e.g. {"table_3": "..."})

steps struct

Each element of steps represents one arithmetic operation:

Subfield	Type	Description
op	string	Operation name with positional suffix (e.g. minus2-1, divide1-2)
arg1	string	First argument: a literal value or #N reference
arg2	string	Second argument: a literal value, const_*, or #N reference
res	string	Result of this step (intermediate or final)

Programs range from 1 to 5 steps. const_100, const_1000, etc. are predefined constants.

Usage

from datasets import load_dataset

ds = load_dataset("rootsautomation/FinQA")
ex = ds["train"][0]

print(ex["question"])
print(ex["answer"])        # human-readable
print(ex["exe_ans"])       # raw numeric or yes/no
print(ex["program"])       # flat DSL
print(ex["program_re"])    # nested DSL
print(ex["steps"])         # structured ops

import json
gold = json.loads(ex["gold_inds"])  # {key: evidence_text}

Image collation

This dataset is text-only. The source PDF page images are available in FinTabNet (CDLA-Permissive-1.0). The id field ({TICKER}/{YEAR}/page_{N}.pdf-{idx}) and filename field ({TICKER}/{YEAR}/page_{N}.pdf) provide a direct join key into FinTabNet's file structure, making it possible to construct an image-grounded version of FinQA.

License

QA annotations: MIT.

Underlying table data is sourced from FinTabNet, which is released under CDLA-Permissive-1.0. The original earnings reports are publicly available SEC filings. The FinQA paper (§2) explicitly verified that CDLA-Permissive-1.0 permits creating and publishing additional annotations over FinTabNet data.

Citation

@inproceedings{chen-etal-2021-finqa,
    title     = "{F}in{QA}: 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",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
    year      = "2021",
    publisher = "Association for Computational Linguistics",
    url       = "https://aclanthology.org/2021.emnlp-main.300",
}