BizBench - TAT-QA Extract: leaderboard
The TAT-QA extraction subset of BizBench (Koncel-Kedziorski et al., 2023): 248 questions over financial-report text and tables whose answer is a number stated in the context. Three-shot accuracy in percent (higher is better).
Metric: Accuracy (%; 248 TAT-QA questions answerable by extracting a number from the text or table, 3-shot). Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 2 70B | 94.4 | #4021 |
| 2 | Mixtral 8x7B | 91.1 | #4051 |
| 3 | GPT-4 | 90.3 | #996 |
| 4 | Llama 2 13B | 88.7 | #8311 |
| 5 | Mistral 7B | 87.9 | #7784 |
| 6 | GPT-3.5 | 84.2 | #2430 |
| 7 | Llama 2 7B | 83.1 | #9031 |
| 8 | mpt-30B | 81.5 | #8820 |
| 9 | falcon-40B | 80.6 | #8304 |
| 10 | starcoder | 75 | #9634 |
| 11 | falcon-7B | 62.9 | #9713 |
| 12 | mpt-7B | 62.1 | #9926 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=bizbench-tat-qa-extract · How It Works · Data refreshed daily, snapshot 2026-11-12.