[arXiv]score: 0.20

LEDGER: 4,999 Annual Reports and 118K Questions for Long-Context Financial LLM Eval

June 12, 2026

LEDGER provides 4,999 full corporate annual reports with figures and tables, labeled with 31 financial KPIs linked to earnings-date market reactions, yielding three benchmarks including a 118,048-question TREC-style KPI retrieval task. It targets long-context extraction capabilities beyond plain-text 10-K filings used in most existing financial NLP benchmarks.

HOW THIS AFFECTS YOU

●

builderYou can use LEDGER to benchmark long-context retrieval and extraction performance of your financial LLM pipelines against a realistic, multi-modal document corpus rather than simplified SEC text.

●

researcherThe KPI-to-market-reaction linkage enables evaluation of whether extracted financial signals carry predictive validity, adding an extrinsic grounding dimension absent from most document QA benchmarks.

read original ↗arxiv.org

← back to feed