foxmedik/GameBoyGhost-LADX

GameBoyGhost-LADX

A dataset of 16,777,216 emulator actions across 8,133 episodes, collected for the GameBoyGhost research project in Link's Awakening DX. Source and experiment documentation: GameBoyGhost.

Contents

• —raw/: 64 Parquet shards, retaining full episodes and all committed rows. Each Parquet row group contains one complete source episode.
• —metadata/episodes.parquet: episode provenance, source checksums, shard and row-group lookup, original curation split/cohort and duplicate relationships.
• —metadata/observation-layout.json: exact byte layout of structured observations.
• —curated/train and curated/dev: clean sword imitation, hindsight navigation, damage events, recovery and failure indexes referencing the raw rows.
• —manifest.json and checksums.json: release details and file integrity.

Rows include physical actions, structured before/after observations, sampled PNG screenshots, room/position/health, sword state, reward components, timing, termination flags, skill and supervision labels. Observation columns are packed binary arrays; decode them with the published layout and the project's gameboy_agent.dataset.unpack_observation helper. Actions encode movement (0 none, 1 up, 2 down, 3 left, 4 right) and buttons (0 none, 1 A, 2 B).

Producer identifiers are pseudonymized. All other raw row values are preserved. ROM files, savestates, reference checkouts, local paths and the reserved evaluation specifications are not included.

Collection and supervision

The fixed sword policy is followed by scripted novelty exploration, with varied physical setup waits and movement perturbations. Collection uses the inherited privilege-D assisted harness, including structured game state and upstream assistance. It is not a neutral pixels-only dataset or a set of human expert trajectories. Full-game completion was not evaluated.

There are 5,348 death outcomes, 1,026 sword timeouts and 1,759 budget-exhausted episodes. Death-ending episodes can contain useful earlier behavior: curation retains safe prefixes and separately labels damage/recovery/failure contexts. Raw data retains 672 duplicate episodes. The curated indexes deduplicate source episodes and contain 321,083 overlapping segments, not 321,083 independent runs.

Only the clean sword lane is imitation-eligible by default. Navigation and recovery are observed-outcome candidates; a successful trajectory can still contain inefficient or conflicting action labels. Users should not blindly clone every action in a successful or failed episode.

Splits and evaluation

The raw split column says train because it records the original collection partition. It is not the subsequent model-training/development assignment. Use episode metadata and curated cohort assignments for that purpose.

The original curation used grouped train/development splits. Some development cohorts were subsequently used for corrective training. These are flagged development_retired=true in episode metadata and curated indexes. Do not use those rows as validation for the published correction experiments. Even the remaining development data shares prepared base states; it is not an independent generalization benchmark. The 96 reserved evaluation specifications were not collected or included in this release.

Loading

from datasets import load_dataset

# Stream the raw collection without materializing all observations in RAM.
rows = load_dataset("foxmedik/GameBoyGhost-LADX", "raw",
                    split="collection", streaming=True)
row = next(iter(rows))

To resolve a curated segment locally, read its release_shard at release_row_group, then slice step_start:step_end within that episode. source_path and source_sha256 retain the pre-release episode provenance. The source checksum differs from the consolidated release-shard checksum.

Verification and limitations

The original collection passed full source/schema checks and sampled exact emulator replay. Every released raw row group was additionally compared value-for-value against its committed source episode after producer-ID pseudonymization. File checksums cover the complete release.

This initial release contains the original collection and its historical curated indexes. Later focused correction demonstrations and model weights are separate local artifacts and are not included here.

Game content remains attributable to its respective owners. No license grant for underlying game assets or third-party material is asserted by this card.