BizBench - ConvFinQA Extract: leaderboard
The ConvFinQA extraction subset of BizBench (Koncel-Kedziorski et al., 2023): 916 questions over earnings-report text and tables whose answer is a number stated in the context. Three-shot accuracy in percent (higher is better).
Metric: Accuracy (%; 916 ConvFinQA questions answerable by extracting a number from the report text or table, 3-shot). Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 | 94 | #996 |
| 2 | Llama 2 70B | 92.8 | #4021 |
| 3 | Mixtral 8x7B | 92.8 | #4051 |
| 4 | GPT-3.5 | 92.4 | #2430 |
| 5 | Mistral 7B | 91.5 | #7784 |
| 6 | Llama 2 13B | 88.2 | #8311 |
| 7 | Llama 2 7B | 86.1 | #9031 |
| 8 | mpt-30B | 85.5 | #8820 |
| 9 | falcon-40B | 82.8 | #8304 |
| 10 | starcoder | 79.7 | #9634 |
| 11 | mpt-7B | 71.7 | #9926 |
| 12 | falcon-7B | 66.5 | #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-convfinqa-extract · How It Works · Data refreshed daily, snapshot 2026-11-12.