Datasetpublic

artefactory/ledger-long-context-multi-kpi

the LEDGER Long-Context Multi-KPI extraction datasets and benchmarks. OCR'd annual reports with ground-truth KPI values for financial information extraction benchmarking. Dataset Description This dataset pairs OCR-extracted annual report text (from DeepSeek OCR) with structured KPI ground-truth values. It is designed for evaluating LLM-based financial information extraction, retrieval, and needle-in-a-haystack tasks. Configs Config Reports… See the full description on the dataset page: https://huggingface.co/datasets/artefactory/ledger-long-context-multi-kpi.

sourceHugging Facecc-by-4.0updated 2mo agoView on Hugging Face

12likes3kdownloads

Dataset Card

the LEDGER Long-Context Multi-KPI extraction datasets and benchmarks.

OCR'd annual reports with ground-truth KPI values for financial information extraction benchmarking.

Dataset Description

This dataset pairs OCR-extracted annual report text (from DeepSeek OCR) with structured KPI ground-truth values. It is designed for evaluating LLM-based financial information extraction, retrieval, and needle-in-a-haystack tasks.

Configs

Config	Reports	Companies	KPI rows	Years	Purpose
`no_eval`	4,505	725	104,529	2009–2024	Training / development
`eval`	494	111	13,519	2017–2022	Benchmark evaluation

Schema

Each row in the parquet files contains:

Column	Type	Description
`ticker`	string	Stock ticker symbol
`exchange`	string	Stock exchange (NYSE, NASDAQ, LSE, AMEX, ASX, OTC)
`company_name`	string	Company long name
`industry`	string	Industry classification
`year`	int	Fiscal year
`revenue`	float64	Total revenue
`net_income`	float64	Net income
`total_assets`	float64	Total assets
`total_liabilities`	float64	Total liabilities
...	float64	31 KPI columns total (see below)
`mmd_text`	string	Full OCR text of the annual report (Markdown with page splits)

KPI columns (31): accounts_payable, accounts_receivable, capex, cash_and_equivalents, cash_incl_restricted, cost_of_revenue, depreciation_amortization, dividends_paid, eps_basic, eps_diluted, financing_cash_flow, gross_profit, income_tax_expense, interest_expense, inventory, investing_cash_flow, long_term_debt_current, long_term_debt_noncurrent, long_term_debt_total, net_income, operating_cash_flow, operating_income, rd_expense, revenue, sga_expense, shares_outstanding, short_term_borrowings, stockholders_equity, stockholders_equity_incl_nci, total_assets, total_liabilities.

KPI values are in millions (as-reported, no FX conversion). NaN means the KPI was not available for that report/year.

Additional Files

—no_eval/mmd/ and eval/mmd/: Raw .mmd files (same text as the mmd_text column, for direct file access).
—eval/images/: Page-level JPEG images for eval reports (not included for no_eval to save space).

OCR Format

The .mmd files use Markdown with page boundaries marked by <--- Page Split --->. Images are referenced as ![](images/{page}_{idx}.jpg).

Usage

python

from datasets import load_dataset

# Load training set
ds = load_dataset("artefactory/ledger-long-context-multi-kpi", "no_eval")

# Load eval set
ds_eval = load_dataset("artefactory/ledger-long-context-multi-kpi", "eval")

# Example: filter to reports with revenue data
ds_filtered = ds["train"].filter(lambda x: x["revenue"] is not None)

Data Sources

—OCR text: DeepSeek OCR applied to annual report PDFs from SEC EDGAR, LSE, ASX, and other exchanges.
—KPI values: SEC EDGAR (XBRL companyfacts) for US listings; yfinance for non-US; Alpha Vantage for gap-fill.

Citation key

If you use LEDGER datasets and / or code ressources, please consider citing our work with:

@misc{moslonka2026ledgerlongcontextbenchmarkcorporate,
      title={LEDGER: A Long-Context Benchmark of Corporate Annual Reports for Grounded Financial Retrieval and Extraction}, 
      author={Charles Moslonka and Amaury de Vitry and Arthur Garnier and Hicham Randrianarivo and Emmanuel Malherbe},
      year={2026},
      eprint={2606.13100},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.13100}, 
}