BizBench - CodeTAT-QA: leaderboard
CodeTAT-QA, a BizBench task (Koncel-Kedziorski et al., 2023): 392 TAT-QA questions answered by writing Python that reads a financial table as a pandas dataframe, correct within 1% of the reference. Three-shot accuracy in percent (higher is better).
Metric: Accuracy (%; 392 TAT-QA table questions answered by generating Python over a dataframe, answer within 1% of the reference, 3-shot). Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 | 90.6 | #996 |
| 2 | GPT-3.5 | 87.6 | #2430 |
| 3 | Mixtral 8x7B | 83.9 | #4051 |
| 4 | Llama 2 70B | 79.1 | #4021 |
| 5 | Mistral 7B | 75 | #7784 |
| 6 | starcoder | 70.2 | #9634 |
| 7 | Llama 2 13B | 65.1 | #8311 |
| 8 | mpt-30B | 64.8 | #8820 |
| 9 | falcon-40B | 38.5 | #8304 |
| 10 | Llama 2 7B | 37 | #9031 |
| 11 | mpt-7B | 30.4 | #9926 |
| 12 | falcon-7B | 7.4 | #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-codetat-qa · How It Works · Data refreshed daily, snapshot 2026-11-12.