LEDGER: A Long-Context Benchmark of Corporate Annual Reports for Grounded Financial Retrieval and Extraction

The LEDGER dataset has been released, comprising 4,999 digitized corporate annual reports that include figures, tables, and narratives, designed for rigorous evaluation of long-context capabilities in financial reporting. It features 31 labeled financial KPIs and provides three evaluation benchmarks: a page-level KPI retrieval task, a conversational single-value lookup, and a full KPI extraction task, all supported by human OCR-quality annotations. This resource is significant for practitioners as it enables the development and benchmarking of models for grounded financial retrieval and extraction, addressing the complexities of interpreting long, numerically dense documents.