BIG-Bench Hard (BBH) Data Quality Investigation Report
收藏Zenodo2025-12-30 更新2026-06-05 收录
下载链接：
https://zenodo.org/doi/10.5281/zenodo.18091194
下载链接
链接失效反馈官方服务：
资源简介：
49 ground truth labeling errors identified in the BIG-Bench Hard (BBH) benchmark dataset through systematic multi-model cross-verification.
Error Summary
date_understanding: 7 errors
geometric_shapes: 42 errors
22 cases: (K) ellipse → (A) circle
20 cases: (K) trapezoid → (H) rectangle
Verification
4-phase verification using GPT-4o, GPT-5.2 Pro, Claude Sonnet/Opus, DeepSeek-Chat/Reasoner, and Qwen3-8B.
Impact
16.8% error rate in geometric_shapes ground truth.
提供机构：
Zenodo创建时间：
2025-12-30
