DataClawEval - HiveSQL: leaderboard
DataClawEval gives agents industrial data-engineering tasks rebuilt from production workflows: they must inspect the data, write and debug a job, and materialize correct output tables. This board covers the tasks that run on HiveSQL. Grading is deterministic, comparing the produced artifact row by row against live databases and adding a process score for exploration, execution efficiency and self-verification.
Metric: Rule-based task score (%; per task a weighted sum, usually 0.7 and 0.3, of the artifact score, checked row by row against live databases by task-specific deterministic scripts, and the process score for exploration, execution efficiency and self-verification, mean over the 28 HiveSQL tasks; one run per task, every model in the same Tencent CodeBuddy agent scaffold). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | CodeBuddy + GPT 5.5 | 69.8 | |
| 2 | CodeBuddy + GPT 5.3 Codex | 67.1 | |
| 3 | CodeBuddy + Claude Opus 4.8 | 65 | |
| 4 | CodeBuddy + DeepSeek V4 Flash | 64.1 | |
| 5 | CodeBuddy + Gemini 3.1 Pro | 63.6 | |
| 6 | CodeBuddy + MiniMax M2.7 | 63.6 | |
| 7 | CodeBuddy + Gemini 3.5 Flash | 62.9 | |
| 8 | CodeBuddy + Hy3 | 62.8 | |
| 9 | CodeBuddy + MiniMax M3 | 62 | |
| 10 | CodeBuddy + Kimi K2.6 | 61.5 |
Interactive version: theaggregate.ai/benchmark?slug=dataclaweval-hivesql · How It Works · Data refreshed daily, snapshot 2026-11-13.