LongGameReasoning

Long-horizon game-playing trajectories with image observations, structured decisions, and tool calls. This repository combines two validated English SFT releases into one dataset. The training splits are assigned by whole game, with no game shared between splits.

Layout

• sft/train.jsonl: 10,396 training records from 17 games.
• sft/dev.jsonl: 1,813 validation records from 6 games.
• sft/test.jsonl: 3,299 test records from 6 games.
• sft/images/: 20,940 deduplicated images referenced by those records.
• sft/images.jsonl and sft/manifest.json: image and release checksums, counts, and provenance.
• raw/base_primary_sources.zip: source annotation directories and screenshots for the original release.
• raw/additional_source_data.zip: original archive supplied for the additional 13 games. This is a preservation copy and includes drafts and work files, not just final training records.
• raw/videos/: 30 videos from the original annotation workspace plus 8 full-length source videos found in the local annotation projects. Some videos are outside the final 29-game SFT selection.
• raw/staging_manifest.json and raw/source_video_manifest.json: archive and video checksums.
• raw/remaining/data2/: 8 ZIP archives of the remaining DATA2 engineering releases, earlier deliveries, review work, and loose annotation material. The 8 original primary-source directories and root-level videos are already preserved separately above.
• raw/remaining/codex/: 12 ZIP archives of local game-annotation projects. These preserve project screenshots, annotations, review outputs, clips, and normalized working videos. Original full-length videos already in raw/videos/ are not duplicated here.
• raw/related_projects/: separate archives of game_agent_dataset (FH_FALL, MACHINARIUM, WIRE_PUZZLE) and game_video_dataset. These projects are not included in the 29-game SFT split.
• raw/training_versions/: historical training-data variants and reports from game-reasoning-sft. These are archival work products, not the validated merged release.

The raw/ files are archival sources. For training, use sft/{train,dev,test}.jsonl; image paths in these records are relative to sft/. Each record has sample_id, split, episode_id, step_id, messages, images, tools, and metadata. The messages contain the observation, structured decision, and tool action. The tools field is a JSON-encoded tool schema string.

The merged SFT release has 15,508 records across 29 games. It excludes 3,193 source records that did not pass the SFT-ready selection. GRPO experiment screenshots are not included.

Rights and limitations

The dataset owner confirmed authorization to publish the game footage and screenshots. This card does not grant a separate downstream license to the underlying games or video content. Verify the relevant rights before redistributing or commercially reusing those assets. The raw source archive may contain work-in-progress material; use the validated SFT split for model training.

The source records preserve some local source-path strings in metadata.source_file for provenance. Image references themselves are portable paths within this repository. Normalized videos and short review clips are in the project archives, not raw/videos/. Browser profiles, cookies, access credentials, software environments, and regenerable package caches were excluded from the public archives. GRPO experiment artifacts and state screenshots were also excluded; they are not backed up here.