Datasetpublic

machinaExMachina/duet-pcg-v1

MultiWorld Duet PCG v1 2-player procedural-generation Minecraft gameplay dataset. Companion to duet-oov-scripted-v1 with added world + task diversity for the MultiWorld paper. Overview Episodes: ~1,700 (3,400+ player streams) Format: Solaris-compatible sharded WebDataset TARs Resolution: 1280×720 @ 20 FPS World diversity: 20 config combos × mixed flat/normal worlds × 5 biomes (plains, snowy_plains, desert, badlands, mushroom_fields) Task diversity: 10 episode… See the full description on the dataset page: https://huggingface.co/datasets/machinaExMachina/duet-pcg-v1.

sourceHugging Facecc-by-nc-4.0updated 5mo agoView on Hugging Face

0likes41downloads

Dataset Card

MultiWorld Duet PCG v1

2-player procedural-generation Minecraft gameplay dataset. Companion to duet-oov-scripted-v1 with added world + task diversity for the MultiWorld paper.

Overview

—Episodes: ~1,700 (3,400+ player streams)
—Format: Solaris-compatible sharded WebDataset TARs
—Resolution: 1280×720 @ 20 FPS
—World diversity: 20 config combos × mixed flat/normal worlds × 5 biomes (plains, snowyplains, desert, badlands, mushroomfields)
—Task diversity: 10 episode types (4 new PCG templates + 6 Phase 1 OOV scenarios)
—Server: Minecraft 1.21 (Paper)

Why this dataset exists

Phase 1 (duet-oov-scripted-v1) established the core OOV scenarios on flat plains worlds. Phase 2 adds world and task diversity: scenarios run across different biomes (including non-flat vanilla terrain) with procedurally varied parameters. The goal is scenario breadth, not raw frame count — per the MultiWorld data plan, "scenario diversity, not raw frame count, drives generalization for the consistency story."

New PCG Task Templates

Four new task-template episode handlers added in Phase 2. Each has procedural parameters drawn from a shared per-episode PRNG seed so both bots make consistent decisions independently.

Scenario	Description	OOV Type
gatherScatter	Separate to regions (40–120 b), mine 3–8 blocks each, rejoin	Decoupled resource gathering
exploreReunite	Explore opposite directions with 2–4 look-around stops, hold OOV, return	Terrain exploration
mobHuntReunite	Spawn mobs at two distant sites (30–80 b), each bot interacts with own mobs, reunite	Mob-interaction asymmetry
buildExchange	Scatter + build (tower/wall/platform/staircase, size 2–5), swap to inspect other's build	Structural variety

Phase 1 Scenarios Also Included

All six Phase 1 OOV scenarios run in Phase 2 with the added world diversity: splitAndRejoin, throughTheWall, farApartBuild, asymmetricInformation, jointWitnessReunite, sharedBuildReunite.

World Configurations

The PCG manifest (pcg/production_manifest.json) defines 20 configurations, each with:

—World seed — deterministic from config index via MD5 salt
—World type — flat (8 configs) or normal (12 configs, vanilla biome generation)
—Biome (flat worlds) — one of plains, snowyplains, desert, badlands, mushroomfields
—Episode type mix — random 3–6 types sampled from the 10-scenario pool

File structure

Same as duet-oov-scripted-v1:

train/<prefix>_<episode>_Alpha_instance_<i>.mp4
train/<prefix>_<episode>_Alpha_instance_<i>.json
train/<prefix>_<episode>_Alpha_instance_<i>_episode_info.json
train/<prefix>_<episode>_Bravo_instance_<i>.mp4
train/<prefix>_<episode>_Bravo_instance_<i>.json
train/<prefix>_<episode>_Bravo_instance_<i>_episode_info.json

episode_info.json includes per-scenario eval_metadata (distances, durations, structure types, mob types, etc.) drawn from the episode's PRNG plan.

Usage

Compatible with the Solaris training pipeline. Extract shards:

bash

for f in *.tar; do tar xf "$f"; done

Citation

If you use this dataset, please cite the MultiWorld paper (forthcoming).