violetxi/ch-trajectory-pool-qwen3.5-4b

C&H Trajectory Pool — Qwen3.5-4B (on-policy)

1,376 agentic exploration trajectories over the full Calderwood & Harkness (C&H) synthetic law-firm corpus (266 matters, ~145M tokens; the open-sourced world from harvey-labs tasks/firm-knowledge/, MIT), generated by Qwen/Qwen3.5-4B — the same model intended as the training student, so this pool is exactly on-policy for it. Part of a world-internalization research project: which likelihood targets, derived from agent experience, internalize a world into model weights.

Companion datasets under this namespace: ch-pilot-rollouts-qwen3.5-35b-a3b and ch-pilot-rollouts-qwen3.5-9b (20-rollout actor-selection pilot).

Task mix (1,376 = one identical, seed-matched spec set)

• —B1 (266) — one survey per matter: client, parties, timeline, events, status, key documents; path citations required for every claim.
• —B1C (46) — one survey per client: cross-matter dossier.
• —B2 (1,064) — self-QA, 4 per matter: given a 2,000-char excerpt from a sampled seed document, pose 3 nontrivial questions (single-doc / cross-doc / cross-matter) and answer them with tools; answers must cite at least one non-seed file, and the cross-matter answer files from ≥2 matters.

Protocol

Text-based tools (glob / grep / read) over a markdown mirror of the corpus, one call per turn as <tool>{json}</tool>, observations returned in <tool_result> (16,000 chars max), answers in <final>...</final>, temperature 0.7, ≤20 turns, 65,536-token context. Generation was mechanically gated (this matters for data quality):

• —a </tool> stop-sequence ends each turn at the tool call (no runaway generation);
• —a <final> is accepted only if closed, ≥50 chars, and after ≥1 successful tool call — otherwise the driver returns a corrective error (3 strikes → protocol_failure);
• —near the context budget (~50k tokens) or turn cap, tool access ends and the model is told to write its final from what it already read (graceful wrap-up instead of overflow loss).

Composition

Each row: full message history (messages: system / user / assistant / tool), run metadata (mode, status, turn/call/error counts, token usage, seed_ref), and a precomputed `usable` flag. Assistant turns include the model's reasoning text.

Provenance & license

All document content is synthetic (no real firm, clients, or people), derived from Harvey AI's MIT-licensed harvey-labs release (commit 7be41d5); trajectories released under MIT. The 250 firm-knowledge eval tasks are not used or contained here — they remain a held-out evaluation set.