BizBench - SEC-Num: leaderboard
SEC-Num, a BizBench task (Koncel-Kedziorski et al., 2023): given a snippet of a 10-K or 10-Q filing and an XBRL label, return the quantity the filer tagged with that label; 2,000 test items from 202 filings. Three-shot accuracy in percent (higher is better).
Metric: Accuracy (%; 2,000 SEC 10-K and 10-Q filing snippets, extract the quantity matching an XBRL label, 3-shot). Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 | 79.3 | #996 |
| 2 | GPT-3.5 | 76 | #2430 |
| 3 | Llama 2 70B | 74.7 | #4021 |
| 4 | Mixtral 8x7B | 73.1 | #4051 |
| 5 | Mistral 7B | 65.1 | #7784 |
| 6 | Llama 2 13B | 61.4 | #8311 |
| 7 | starcoder | 57.8 | #9634 |
| 8 | mpt-30B | 56.4 | #8820 |
| 9 | Llama 2 7B | 56.2 | #9031 |
| 10 | falcon-40B | 52.3 | #8304 |
| 11 | mpt-7B | 31.7 | #9926 |
| 12 | falcon-7B | 27.3 | #9713 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=bizbench-sec-num · How It Works · Data refreshed daily, snapshot 2026-11-12.