Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
The size of the content of the first rows (8351234 B) exceeds the maximum supported size (200000 B) even after truncation. Please report the issue.
Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

PPI TextWorld Games

Immutable, precompiled parser-game bundles for the Friday & Co PPI/verl TextWorld agent loop. Rollout jobs download one archive at an exact Hugging Face commit and verify the archive, manifest, every Z-machine ROM, and every TextWorld JSON sidecar before use. Games are generated and compiled offline; rollout workers never generate tasks.

Procedural v2

ppi-textworld-procedural-v2.tar.gz contains 6,000 unique solvable games with frozen 5,000/500/500 train/validation/test membership. Every selected game has a fresh successful oracle replay, and the release rejects duplicate game IDs, ROM hashes, sidecar hashes, or structural hashes globally across all splits.

Family Total Train Structural difficulty control
native_chain 2,520 2,100 native quest length and breadth
fetch_delivery 1,800 1,500 rooms, source state, destination type, distractors
nested_key_delivery 600 500 nested lock depth
multi_delivery 600 500 number of targets and destinations
secure_delivery 480 400 route scale and locked delivery chain

The native family exercises TextWorld's procedural quest generator. The four custom families construct maps, rooms, named objects, containers, supporters, keys, locks, multi-object goals, and explicit win conditions through TextWorld primitives. Component seeds for map, objects, quest, and grammar are deterministic and recorded per game.

The v2 release ships these reviewable provenance artifacts alongside the archive:

  • ppi-textworld-procedural-v2.manifest.json: immutable split membership, relative paths, hashes, family/settings/difficulty, component seeds, structural metadata, and oracle length for every game;
  • textworld_procedural_v2_release_plan.json: exact 21-cell mixture and calibration evidence used to select it;
  • textworld-procedural-calibration-2026-08-06.md: model wire, per-cell results, posterior calculation, full 5,000-game results, anomaly inspection, contamination checks, and latency;
  • textworld_procedural_v2_full_audit.json: machine-readable per-trajectory results and integrity, lifecycle, concurrency, latency, and model-output anomaly summaries;
  • build, failure-journal, and independent-verification summaries for the compiled corpus.

The release candidate was calibrated with friday-and-co/Qwen3.5-2B@d04857b5fec68b427d7504c8740b674da15d3ff7 using thinking mode and the qwen3_coder tool-call format. Its frozen mixture has a 17.82% weighted empirical point estimate and 18.47% Jeffreys-posterior median. A fresh rollout over all 5,000 training games observed 935 wins, or 18.70% pass@1 (95% Wilson interval [17.64%, 19.80%]). All 5,000 trajectories passed the reward, score, identity, lifecycle, command-trace, and cross-rollout contamination checks.

Legacy v1

ppi-textworld-games-v1.tar.gz remains available unchanged. It contains 336 games from TextWorld's built-in coin_collector, cooking, and treasure_hunter challenges, with frozen 192/48/96 train/validation/test membership. Existing immutable revision pins remain valid; v2 is a separate artifact rather than an in-place replacement.

Runtime and upstream notices

Both bundles target TextWorld 1.7.0 and Jericho 3.3.1. Archives intentionally exclude generated Inform 7 (.ni) intermediates because rollouts need only .z8 ROMs and JSON sidecars.

TextWorld is distributed under the MIT license; its documentation identifies Inform 7 under the Artistic License 2.0 and Jericho under GPL-2.0. Those upstream projects and licenses remain authoritative. This dataset card does not relicense third-party components.

Source project: https://github.com/microsoft/TextWorld

Downloads last month
148