xixibuxixi/polis-v1

Polis v1 (360p)

Polis v1 is a synchronized multi-resident game dataset for learning a writable 3D world model, joint state transitions, video rendering, and resident policies. It contains human-like players and NPCs engaged in construction, movement, cooperation, PvE, and PvP. Episodes are stored as continuous trajectories rather than model-specific fixed-length windows.

Complete reproduction download

This dataset repository is also the single download entry point for the project. The continuous episode shards remain under data/. Versioned archives under bundles/ contain the matching Polis source, derived training indexes, environment specifications, required Polis checkpoints, and the packaged Gamma-World and Solaris comparison code/checkpoints.

huggingface-cli download xixibuxixi/polis-v1 \
  --repo-type dataset --local-dir polis-v1

Users who already have the episode shards can download only the reproduction archive and its manifest from bundles/. A separately downloadable, tested Linux x86-64 virtual environment is also available there for quick startup. It uses Python 3.11.14 and PyTorch 2.7.1 with CUDA 12.8. Dependency lock files and installation recipes remain available for other platforms and CUDA stacks.

This compact release materializes RGB video and lossless uint16 instance masks at 640 x 360. It was derived from the retained 1280 x 720 source using Lanczos filtering for RGB and nearest-neighbor index selection for instance IDs.

Scale

Split	Episodes	Frames
train	40,403	13,679,582
val_id	2,576	890,755
test_id	2,593	907,772
total	45,572	15,478,109

The release contains 68,782 damage events, 70,584 attack contacts, 93,475 attack attempts, and 321,667 block-edit events. The training split uses the deterministic balanced_v1 selection; validation and test are retained in full. See release.json for scenario/activity counts and the exact selection policy.

Distribution format

The Hub copy groups complete episode directories into uncompressed tar shards under data/{split}/. Tar preserves hard links used for compatibility aliases, avoids more than one million individual Hub files, and supports sequential or selective extraction. shards.jsonl records each shard's byte size, SHA-256, and episode count after the upload completes.

The top-level manifests are also published directly:

• train.jsonl, val_id.jsonl, and test_id.jsonl: one searchable record per episode;
• episodes.jsonl: the combined episode catalog;
• balanced_v1_train.jsonl: the selected training catalog;
• release.json: aggregate statistics and balancing policy;
• COMPLETE.json and DERIVATION.json: completion and spatial provenance.

Episode contents

Each episode directory includes synchronized payloads such as:

• data.npz: voxel observations, player/camera state, actions, inventory, entities, health, and uint16 instance masks;
• rgb_agent*.mp4: one 360p RGB stream per resident viewpoint;
• trajectory.jsonl, events.jsonl, and protocol.jsonl: continuous state and event records;
• manifest.json, training_metadata.json, summary.json, and validation.json: task definition, schema, outcome, and quality checks;
• m1_initial.npz and m1_rgb_agent*.jpg: first-frame initialization payloads;
• world_replay.json: initial world state needed for replay.

For an episode with T actions and N residents, observations have length T + 1. Major arrays include obs_voxel_mt [T+1,N,49,49,49,2], obs_voxel_center [T+1,N,3], player and camera state, action_keys [T,N,21], continuous mouse/action arrays, entity tracks, health supervision, and instance_mask [T+1,N,360,640]. The mask is supervision rather than a model condition. ID 0 is background and 65535 denotes the observer's wielded item; the remaining active IDs map through the per-frame entity tables.

Natural health regeneration is disabled in combat data. Attack attempt, contact, and damage are recorded separately.

Reading a shard

huggingface-cli download xixibuxixi/polis-v1 \
  data/train/train-00000-of-00082.tar --repo-type dataset --local-dir polis-v1
tar -xf polis-v1/data/train/train-00000-of-00082.tar

The train, validation, and test splits contain 82, 6, and 6 shards respectively. The tar member paths retain the split prefix, for example train/<episode_id>/data.npz.

Validation and intended use

All included episodes passed the release's success && validation.usable filter. The dataset is intended for research on learned simulation, embodied agents, multi-agent dynamics, 3D-aware rendering, and structured policy models. Users should preserve episode boundaries and action/observation alignment when constructing training windows.