README.md

metadata

language:
  - en
tags:
  - mctd
  - on-policy-distillation
  - agent-traces

MCTD optional OPD annotation snapshots

Consistent SQLite backups of the existing ALFWorld, ScienceWorld, PlanCraft and EB-Habitat teacher-output stores. They retain outputs regardless of success, with separate trajectory_outcomes labels where available. Missing outcome labels mean unknown, not success. Publication does not re-label these entries.

Environment	Output records	Outcome records
ALFWorld	275,140	243,700
ScienceWorld	15,246	15,246
PlanCraft (2026-09-10 snapshot)	16,197	11,015
EB-Habitat	39,938	0

The latest PlanCraft pack also contains 1,750 teacher-results cache rows. The main output archive includes 403 additional unique outputs from the earlier Qwen3.8 archive, preserving its 432 additional outcome associations. It installs immutable databases under data/opd_cache/plancraft/snapshots/20260910/; the OPD workflow seeds separate writable runtime stores on first use. Existing live caches on the source server are reused without being overwritten by a download. The older PlanCraft pack remains available in repository history and as a versioned artifact; the current manifest selects the newer snapshot.

Outputs and outcomes are not necessarily one-to-one. The same context may have multiple results and episode outcomes. SQLite quick_check, SHA-256 and archive inventories are checked during export. Use the exact manifest counts if a newer revision differs from this card.

python scripts/download_resources.py --environment all --with-cache

Reuse contract

These caches preserve original namespace/context/payload keys, provenance, quality flags and outcomes. Historical path or model-contract differences can prevent cache hits on another machine. Only compatible entries may be reused; this snapshot does not promise that every historical output will hit, and the download does not rekey incompatible E+C supervision.

Duplicate static-success indexes are deliberately omitted: the existing training code rebuilds them from the teacher dataset in oosnow/mctd-data. These caches are optional acceleration inputs, not success-only SFT data. Keep the algorithm's outcome-aware reuse and reward-weighting rules.

Model management labels are not proof of the actual generator. Preserve the original backend provenance; multiple teacher and scorer runs are represented. The files contain no model weights, API credentials or training optimizer state. Environment-derived/generated material remains subject to applicable upstream/provider terms; this card does not establish a new blanket license.

The Habitat snapshot (eb-habitat-20260909.tar.gz) has no separate trajectory outcome records. These entries are unknown-outcome outputs, not successful or failed trajectory labels. They must not bypass outcome-aware reuse rules. Successful static trajectories are distributed separately in mctd-data.