siddharthmb/

collab-arena-v0-transcripts

⚠️ This dataset is LIVE and GROWING until the experiment completes. Configs and counts update as Qwen cells and the 32B tier land; see the commit history for the update log.

ℹ️ E5 frontier model = claude-opus-4-8 (E1-E4 frontier = claude-fable-5). Fable safety-refused ~50% of E5 turns (stop_reason=refusal), concentrated on the honest cross-check seats, so E5's frontier arm was run on Opus instead; team and solo are BOTH Opus… See the full description on the dataset page: https://huggingface.co/datasets/siddharthmb/collab-arena-v0-transcripts.