LEDGER - Conversational KPI Lookup: leaderboard
LEDGER (2026) is a corpus of 4,999 digitized corporate annual reports labelled with 31 financial KPIs. On this board the model reads a whole OCR-extracted test report of about 100k tokens and answers one of 10,000 single-KPI questions over 494 test reports; the score is recall, the percent of answers within 0.05 percent of the labelled value. Higher is better.
Metric: Recall (%) on 10,000 single-KPI questions over the 494 test annual reports: the model reads a whole OCR-extracted report (about 100k tokens) and returns one requested KPI as JSON (found, value, unit scale, page); a value counts when within 0.05% of the label; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.6 27B | 91.4 | #341 |
| 2 | Ministral 3 14B | 87.9 | #1845 |
| 3 | GPT-OSS-20B | 85.3 | #1147 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ledger-conversational-kpi-lookup · How It Works · Data refreshed daily, snapshot 2026-11-13.