BizBench - CodeTAT-QA: leaderboard
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 |
|---|---|---|
| 1 | GPT-4 | 90.6 |
| 2 | GPT-3.5 | 87.6 |
| 3 | Mixtral 8x7B | 83.9 |
| 4 | Llama 2 70B | 79.1 |
| 5 | Mistral 7B | 75 |
| 6 | starcoder | 70.2 |
| 7 | Llama 2 13B | 65.1 |
| 8 | mpt-30B | 64.8 |
| 9 | falcon-40B | 38.5 |
| 10 | Llama 2 7B | 37 |
| 11 | mpt-7B | 30.4 |
| 12 | falcon-7B | 7.4 |
Interactive version: theaggregate.ai/benchmark?slug=bizbench-codetat-qa · How It Works · Data refreshed daily, snapshot 2026-09-26.