C&H note-conditioned trajectories from Qwen3.5-27B
14,688 of 31,000 matched tasks have terminal records. This release is partial; remaining tasks will be added.
Policy: Qwen/Qwen3.5-27B at fc05daec18b0a78c049392ed2e771dde82bdf654. Notes: violetxi/wm-rl-notes-v4 at 6e62edba847d8013a5e7aefb1e7fd19974fa5381. Tasks match the original 20K and unused 11K cohorts of violetxi/harvey-note-conditioned-rollouts.
The default student configuration contains only role/content messages, preserving every assistant reasoning turn and removing supplied note memory. Complete examples longer than 16,384 tokens and structurally incomplete rollouts are excluded, without semantic or task-success filtering. The base actor system prompt is provided in actor-system.txt.
The teacher configuration retains every terminal status, exact chronological user/assistant messages, full_trajectories, action-only action_only_trajectories, the exact system_prompt with supplied notes, and structured provenance, token usage, timing, final answers, and raw event records. Action-only assistants contain parsed commands and their local task_complete decision.
Generation: thinking enabled, temperature 0.6, top-p 0.95, top-k 20, seed 731, at most 20 turns, 65,536 context tokens, and 16,384 output tokens per turn. Tool observations and raw model outputs are verified; task correctness has not been graded. Access requests require manual approval.
- Downloads last month
- 27