README.md

metadata

language:
  - en
tags:
  - mctd
  - embodied-agents
  - alfworld
  - scienceworld
  - plancraft
  - eb-habitat

MCTD training resources

Versioned resource packs for the MCTD github_release training/evaluation code. Extract paths are relative to the release root. Use its python scripts/download_resources.py --environment all --with-pretrained instead of manually moving files. resources.json records SHA-256, sizes, destinations and quality flags for every file. These archives are not an automatic Hugging Face load_dataset table.

Teacher dataset	Successful episodes	Steps
ALFWorld	3,553	37,471
ScienceWorld	874	17,302
PlanCraft	1,145	8,839
EB-Habitat (custom text-only)	1,720	14,989

These are collected and hindsight-supplemented trajectories retained by the project's validated export. Publication verifies JSON parsing, index counts, and success=true, done=true; it does not independently replay each episode. Only use training tasks for gradient updates. The repository also contains student-compatible command tables and PlanCraft official-planner trajectories. Small official resources and evaluation protocols are shipped with the code.

Supervision quality and provenance

Successful task completion does not establish that every reasoning field is correct or complete. Preserve and honor per-step quality/availability flags. Do not treat absent curves as teacher probabilities or invent missing E+C.

Dataset	Hindsight/answer-conditioned steps	Truncated reasoning steps	Known missing E+C steps
ALFWorld	10,149	31	0
ScienceWorld	1,677	96	379
PlanCraft	6,848	114	0

Some legacy eligibility flags remain true on truncated steps (30 ALFWorld, 90 ScienceWorld, 1 PlanCraft); consumers must also check the quality flags. The original manifests and supervision payloads are preserved, not repaired or re-generated during upload. Model tags used for dataset management are not necessarily the actual generator identity. The data combines multiple collection/supplementation runs and model backends; consult generator and provenance fields rather than treating all annotations as a single Qwen run. Intermediate E+C may be student-scored rather than teacher-scored.

External resources and rights

Base model weights are not hosted here. Obtain Qwen3-1.7B and Qwen3-30B-A3B-Instruct-2507 from the official Qwen repositories. ALFWorld raw task runtime archives are downloaded directly from the official ALFWorld releases, not mirrored here. Upstream references: ALFWorld, ScienceWorld, PlanCraft. Environment-derived material remains subject to its upstream terms, and generated content to applicable provider terms. This card does not grant a new blanket license or assert ownership of third-party environments/content.

This snapshot is for reproducible research, not a new benchmark split or an independent accuracy claim. It contains no API credentials or optimizer states.

Habitat extension (2026-09-09)

The Habitat package adds 1,720 successful trajectories across 254 training tasks, the immutable 300-task protocol (254 train / 46 test), the 70-action vocabulary, and verified expert records. This is a custom text-only, outcome-selected protocol, not the official visual EB-Habitat split. Preserve simulator episode mappings. The dataset contains 75 collected and 1,645 hindsight trajectories; original provenance and quality fields are retained. Its manifest reports E+C for all 1,720 episodes, not an independent guarantee of the correctness of each annotation.

The raw official model-rollout archive is available from EmbodiedBench/EB-Habitat_trajectory_dataset at revision d391eba8684b0de0c4f37a223e69237c91444209. Simulator sources and ReplicaCAD/YCB/Fetch assets are fetched from pinned official upstreams by scripts/setup_conda.sh eb-habitat; they are not mirrored here. No failed raw-collection workdir, empty command-table placeholder or credentials are included. Latent SFT constructs its real command table from the installed student once. Use python scripts/download_resources.py --environment eb-habitat.