Contains 244 Cursor agent sessions recorded from Claude Fable‑5, formatted for training and research. Sessions include multi-turn assistant/tool interactions and are Teich-compatible; several rows exceed one million characters, so apply explicit tokenization and oversize policies before training.
Provides aggregated, privacy-preserving cluster outputs from three external research teams' analyses of ~250k Claude/Claude Code conversations; includes per-team CSVs (Stanford, Oxford, METR) for studying human–AI collaboration and model behavior without raw conversations.