Joshyxwa/cp2107-textworld-trajectories

Simulator-Grounded ScienceWorld and ALFWorld Trajectories

This repository contains offline, simulator-grounded trajectories used to study whether language models can represent action-conditioned world transitions. Each trajectory records a goal, the initial observation, and ordered Action/Observation blocks from the environment. The train, validation, and test partitions are kept separate.

Releases

Configuration	Train	Validation	Test	Total	Construction
scienceworld	3,534	1,762	1,799	7,095	Compact trajectories recorded from ScienceWorld 1.2.3
alfworld	3,250	133	130	3,513	Successful deterministic replays of ALFWorld expert solutions

ScienceWorld uses the trajectory-v2 schema. ALFWorld uses the compatible trajectory-v1 block schema and contains 93,651 successful transitions. Its actions were verified as admissible before execution, observations came from the simulator, and every retained trajectory reached its goal. The 314 unsuccessful replay attempts are excluded from the split files and documented separately in alfworld-rollouts-v3/replay-failures.jsonl.

Stable Hub revisions are provided as scienceworld-v2 and alfworld-v3. The latter revision contains both configurations.

Load the data

from datasets import load_dataset

scienceworld = load_dataset(
    "Joshyxwa/cp2107-textworld-trajectories",
    "scienceworld",
    revision="alfworld-v3",
)

alfworld = load_dataset(
    "Joshyxwa/cp2107-textworld-trajectories",
    "alfworld",
    revision="alfworld-v3",
)

Each JSONL record includes:

• trajectory_id, env, split, and goal;
• ordered blocks containing Goal, Action, Observation, and—where available—Think text;
• source and simulator metadata; and
• leakage flags and optional state labels.

Frozen evaluation targets

The ALFWorld release includes deterministic position-stratified manifests in alfworld-rollouts-v3/target-manifests/. They select early, middle, and late next-action and next-observation targets without revealing future blocks. The manifests contain 19,500 training targets, 798 validation targets, and 780 test targets.

Reproducibility and limitations

Both releases include manifests describing their construction. ALFWorld additionally includes file checksums, the replay-failure ledger, source-tree provenance, deterministic expert seeds, and frozen target manifests.

These are offline expert trajectories rather than interactive agent runs. They measure prediction and representation under recorded histories; they do not by themselves establish that a model can complete tasks interactively. The successful-only ALFWorld split also should not be interpreted as the environment's natural success distribution. Use of the underlying environments and source assets remains subject to their respective upstream terms.