marjanmoodi/AutoResearch-ECS-V0

AutoResearch ECS World Model Dataset (V0)

Entity-Component-System (ECS) game state sequences for training world models. 10 classic games in two formats and two sizes.

Dataset Variants

Variant	Episodes	Frames	Format D	Parquet
small	~1K	329K	small/format_d/ (34MB)	small/parquet/ (115MB)
full	~360K	98.4M	full/format_d/ (9GB)	full/parquet/ (7.3GB)

Dataset Structure

{small,full}/
├── format_d/                    # Format D: per-game tar.gz archives
│   ├── {game}.tar.gz            # Extract → game={game}/ep_*/{ *.npy, meta.json }
│   ├── unified_manifest.json    # Cross-game entity/action/global field definitions
│   ├── tensor_config.json       # Build config (velocity, physics material, etc.)
│   └── build_stats.json         # Episode/frame counts per game
│
└── parquet/                     # Parquet: one file per split per game
    └── {game}/
        ├── train.parquet        # ~70% of episodes
        ├── val.parquet          # ~15%
        ├── test.parquet         # ~15%
        └── meta.json            # registry_dim, state_dim, max_entities

Games (full dataset)

Game	Episodes	Max Entities	Frames
asteroids	30,000	21	9.7M
breakout	14,995	52	7.5M
flappy_bird	50,000	10	9.5M
frogger	50,000	27	2.0M
platformer	20,000	24	5.6M
pong	50,000	5	21.4M
snake	50,000	21	4.1M
space_invaders	15,000	56	6.1M
tag	30,000	9	29.9M
tetris	50,000	64	2.6M

Total: 359,995 episodes, 98.4M frames

Tensor Schema

Tensor	Shape	Description
registry	(N, 34)	Static entity properties (collider, scale, physics)
states	(T, N, 23)	Dynamic: pos_xy(2), alive(1), vel_xy(2), gameplay(14), pos_history(4)
actions	(T, 7)	Unified action vector (7 fields across all games)
globals	(T, 17)	Global game state (17 fields across all games)
terminals	(T,)	Episode termination flags
mutable_mask	(N,)	Which entities are prediction targets
type_ids	(N,)	Global entity type IDs
slot_ids	(N,)	Original 64-slot table indices
rewards	(T,)	Per-frame rewards

Usage

Parquet (recommended for training)

from huggingface_hub import hf_hub_download

# Download one game (full dataset)
path = hf_hub_download(
    "marjanmoodi/AutoResearch-ECS-V0",
    "full/parquet/pong/train.parquet",
    repo_type="dataset",
)

import pyarrow.parquet as pq
import numpy as np, json
table = pq.read_table(path)
row = table.to_pydict()
states = np.frombuffer(row["states"][0], dtype=row["states_dtype"][0]).reshape(
    json.loads(row["states_shape"][0])
)

Format D (for custom pipelines)

from huggingface_hub import hf_hub_download
import tarfile

path = hf_hub_download(
    "marjanmoodi/AutoResearch-ECS-V0",
    "full/format_d/pong.tar.gz",
    repo_type="dataset",
)
with tarfile.open(path) as tar:
    tar.extractall("./data")
# Now: ./data/game=pong/ep_*/{ registry.npy, states.npy, ... }

Generation Pipeline

1. ecs-world generate_all.py — simulate games, record JSONL
2. ecs-world FormatDCacheBuilder — JSONL to Format D numpy tensors
3. ecs-vanilla-baselines export_parquet.py — Format D to Parquet with deterministic splits