DataGovBench - Table QA: leaderboard
DataGovBench (Fujitsu Research of America, 2026) tests data analysis on real government open data: 211 simple and decomposable questions over 178 large multi-table datasets with metadata and external knowledge, answered in text or as charts by generated Python code. Accuracy in percent counts a question only when every sub-question is right, with text judged by exact match and charts by a majority of four multimodal judges; higher is better. The model writes code once from the first ten table rows, without an agent framework.
Metric: Question-set accuracy (%; 211 simple and decomposable questions over 178 government open-data datasets, a set counts only when every sub-question is right; text answers by exact match, charts by a majority of four MLLM judges; the model writes Python once from the first 10 table rows, without the Answer Agent). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 33.7 | #161 |
| 2 | Gemini 2.5 Flash | 31 | #453 |
| 3 | GPT-5.1 (2025-11-13) | 28.9 | #187 |
| 4 | GPT-4o (2024-08-06) | 24.2 | #740 |
| 5 | Devstral Small (2507) | 15.2 | #1836 |
| 6 | Qwen 3 30B A3B 2507 Instruct | 13.4 | #1089 |
| 7 | Qwen 3 Coder 30B A3B Instruct | 12.3 | #972 |
| 8 | DeepSeek R1 Distill Qwen 14B | 3.8 | #2582 |
| 9 | Llama 3.1 8B Instruct | 1.9 | #4246 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=datagovbench-table-qa · How It Works · Data refreshed daily, snapshot 2026-11-13.