Episode corpora, failure-mode detection outputs, and fixed-position residual-stream
activations from a study of failure modes in multi-agent fact verification. A
Qwen3-32B "verifier" directs four Qwen3-8B evidence-holding "subagents" over claims
from the AVeriTeC benchmark; episodes vary
injected conditions (trust/source-tier framing, confidence phrasing, sycophancy
challenges) to elicit and measure failure modes such as transmission loss… See the full description on the dataset page:
https://huggingface.co/datasets/siddharthmb/multiagent-verification-failure-modes.
