AceAction/AceAction-5M_GameForPeace

Game for Peace VLA

The primary Game for Peace dataset release from 3a-vla, pairing a current game image with a task instruction, the next keyboard/mouse action, and a short local action description stored as reason.

Repository: 3a-vla/GameForPeace-VLA · Version: 1.1.0 · Configuration: peace_4hz

This release packages the aligned game frames and instruction/action/reason annotations from the 3a-vla SFT workflow, with image cleanup applied in v1.1.0. The five columns are id, image, instruction, action, and reason. Segmentation targets and masks are outside the scope of this release.

Image cleanup in v1.1.0

This version removes detected cyan/blue enemy corner brackets from RGB images and repairs distance-label regions where they can be localized. It follows the locally reviewed cleanup examples. Bracket geometry, colour, and distance-label templates validate the repair regions. Clipped frame-edge strokes receive an additional pass, and nearby frames can confirm weak detections. RGB repair always uses the current frame only.

Modified frames are stored as lossless PNGs. Every decoded pixel outside the repair mask is verified identical to the source frame. Unmodified frames retain their original JPEG bytes. Covered pixels are estimated from local image context with OpenCV Telea inpainting; distance-label regions also use horizontal context interpolation.

103,505 of 2,522,808 distinct source frames were repaired. Sample IDs, instructions, actions, reasons, and train/test membership are unchanged. The detector is heuristic; the checks do not establish that every possible overlay was found. Aggregate checks and the preprocessing recipe are in metadata/cleanup_validation.json and preprocessing/.

This repository replaces the former game_for_peace_vla_4hz_with_reason repository under a shorter name. The previous remote repository and its history were removed at the owner's request; this repository contains the cleaned release only. Local source archives were retained during replacement.

The related step-3000 model was trained on the earlier marked-image pipeline and has not been retrained on this release. Its model card records fresh predictions on the cleaned examples.

At a glance

Property	Value
Game	Game for Peace / 和平精英
Observation	One current 320×224 RGB image
Action interval	0.25 seconds / 4 Hz
Annotation language	Chinese
Train	7,227,144 rows; 1,743 source recordings
Test	1,200 rows; 274 source recordings
Distinct current-frame paths	2,522,808
Storage	Parquet with embedded JPEG / PNG bytes, about 53.69 GB
Files	128 train shards and one test shard

The train and test source recordings are disjoint. The fixed 1200-example test set has also been used during development and model selection.

Load the dataset

pip install "datasets==4.0.0" "pyarrow>=21" "pillow>=10"

If the repository is private, authenticate with hf auth login using an account that has access.

from datasets import load_dataset

train = load_dataset(
    "3a-vla/GameForPeace-VLA",
    name="peace_4hz",
    split="train",
    streaming=True,
)
train = train.shuffle(seed=42, buffer_size=10000)
sample = next(iter(train))

print(sample["instruction"])
print(sample["action"])
print(sample["reason"])
sample["image"].save("preview.png")  # PIL image, already decoded

Rows sharing a current frame are grouped for efficient image dictionary storage. Shuffle the training stream as above when constructing training batches. Full instructions, actions, and reasons are retained for every sample.

Images are embedded in the Parquet shards. Image decoding requires no separate archive extraction or access to paths on the dataset author's machine.

Fields

Column	Type	Meaning
id	string	Original sample ID, normalized to a string. Unique across both splits.
image	Hugging Face Image	Current-frame RGB image; defaults to a PIL image when loaded.
instruction	string	Original overall task instruction.
action.dx	int32	Original horizontal mouse delta for the next 0.25 seconds.
action.dy	int32	Original vertical mouse delta for the next 0.25 seconds.
action.keys	list of strings	Controls held for that action interval. Empty list means no held controls.
reason	string	Original current-frame local instruction / short action description.

Allowed controls, in canonical output order: w, a, s, d, c, z, f, space, tab, L, R. Mouse deltas retain the source integer units: no gain scaling, degree conversion, clipping, averaging, or aggregation is applied.

The original local-instruction annotations are at 1 Hz and are shared by the corresponding four 0.25-second action substeps. The reason field stores that annotation verbatim; its granularity differs from the action frequency. All train reasons are nonempty. The test split retains its 17 empty reason annotations as empty strings.

The Image feature physically stores {"bytes": ..., "path": ...}. Its path is an image content hash plus .jpg or .png; the image bytes are embedded. To inspect bytes instead of decoding them:

from datasets import Image

raw = train.cast_column("image", Image(decode=False))
image_bytes = next(iter(raw))["image"]["bytes"]

Navigation and combat examples

These two examples are from the held-out test set. Their actions and reasons are reference annotations. The JSON examples use repository-relative image paths for readability; the actual Parquet rows embed the image bytes.

Navigation

{
  "id": "73107",
  "image": "examples/navigation.jpg",
  "instruction": "前进进入红砖房室内",
  "action": {
    "dx": 0,
    "dy": 0,
    "keys": [
      "w"
    ]
  },
  "reason": "向前方红白建筑前进，保持对前方红白建筑的观察。"
}

Combat

{
  "id": "23727",
  "image": "examples/combat.png",
  "instruction": "射击敌人",
  "action": {
    "dx": -3,
    "dy": 1,
    "keys": [
      "L"
    ]
  },
  "reason": "开火射击前方46米处的目标。"
}

Optional SFT format

The structured columns can be adapted to different model families. For the existing Qwen3.5 recipe, serialize the assistant target as:

(dx,dy) key1 | key2 <think>current-frame local instruction</think>

Use NONE for an empty held-control list. The local description follows the action. The chat template appends its native end-of-turn marker.

The full reference system prompt is:

Follow the task instruction using the current game image. Predict the action to execute for the next 0.25 seconds. Output (dx,dy) followed by the held control tokens separated by |, then a space and <think>the local instruction for the current frame</think>. dx and dy are integers in the original annotation units. Use NONE when no controls are held. Allowed control tokens, in output order: w, a, s, d, c, z, f, space, tab, L, R. Output the action before the local instruction. Do not add an END marker.

scripts/format_samples.py provides the system prompt, action formatter, target formatter, and Alpaca record adapter. metadata/training_format.json records the complete formatting contract.

To export a small training subset for LLaMA-Factory, download this repository's scripts/ directory and run:

python scripts/export_alpaca.py --split train --limit 100 --output-dir gfp_train

Remove --limit 100 for the full split. The exporter writes JSONL, image files, and a dataset_info.json registry entry. It copies the embedded image bytes and reuses images shared by multiple rows. Point LLaMA-Factory's dataset_dir and media_dir to gfp_train; use qwen3_5_nothink for the reference template and train_on_prompt=false to supervise the assistant response.

For action-only inference with that recipe, stop generation at <think>. Training can still supervise the full action and postposed local description.

Preparation and validation

• Preserved the existing fixed train/test assignment and all sample IDs.
• Parsed and round-trip-checked every original action/reason response.
• Verified the original annotation-shard checksums during export.
• Preserved instruction text, action integers, control sets, and reason text.
• Checked every distinct embedded image's format, dimensions, and RGB mode.
• Repaired detected enemy overlays; preserved every decoded pixel outside repair masks.
• Preserved original bytes for unmodified JPEGs; saved repaired frames as lossless PNGs.
• Verified sample-ID uniqueness and zero recording overlap between splits.
• Grouped shared frames for Parquet compression; training row order changes.

Counts, schema, source annotation hashes, and shard checksums are provided in metadata/release.json and SHA256SUMS.

Provenance and related model

This release derives from the existing 3a-vla Game for Peace single-frame, single-action 4 Hz conversion. Original image-corpus provenance identifies theend1234/long_events_peace_600_320x224_jpg_0930. Source annotation file checksums are recorded in the release metadata.

The related GP action/think/seg model uses the same base image/instruction/action/local-description records plus additional segmentation supervision. That auxiliary supervision is a separate training extension.

License

This release does not currently specify a dataset reuse license.

Citation

@dataset{three_a_vla_game_for_peace_2026,
  author = {{3a-vla}},
  title = {Game for Peace VLA},
  year = {2026},
  version = {1.1.0},
  url = {https://huggingface.co/datasets/3a-vla/GameForPeace-VLA}
}