barissozudogru/

mast-failure-mode-gold

25 human-verified multi-agent failure traces, each labelled with one of the 14 MAST failure modes (Cemri et al., "Why Do Multi-Agent LLM Systems Fail?", arXiv:2503.13657). Built as the ground-truth set for evaluating the MAST classifier in adk-agent-playground — i.e. for Cohen's κ against an LLM judge, not for training.

	What's in a row




	

field meaning

id stable id, e.g. gold-001

pipeline_name / agent_name which agent produced… See the full description on the dataset page: https://huggingface.co/datasets/barissozudogru/mast-failure-mode-gold.