microsoftarchive/promptbench

Provides modular support for prompt engineering techniques (few-shot chain-of-thought, emotion prompting), adversarial robustness evaluation via prompt attacks, and dynamic test data generation to mitigate contamination. Built on PyTorch with extensible components for datasets, models, and evaluation methods, integrating specialized frameworks like DyVal for dynamic evaluation and PromptEval for efficient multi-prompt assessment across standard benchmarks (MMLU, BigBench Hard, GLUE) and multi-modal datasets.

2,787 stars.