MiniGrid DoorKey fairness v1

Custom initial-state holdout splits for MiniGrid-DoorKey-8x8-v0, protocol minigrid-doorkey-8x8-h64-v1, state schema minigrid-semantic-state-v1, policy prompt version minigrid-vlm-v1. This is not an official Farama evaluation split and does not measure generalization to unseen map structures. Train and validation may share wall layouts. No policy/model calls or model performance enter selection.

Frozen membership

• Generation seeds: all integers 0 through 4095, ascending, NumPy 1.26.4.
• Generated states: 4096; duplicates: 753; unreachable within 64: 0.
• Eligible deduplicated frozen candidates: 3343.
• Canonical validation: exactly 256 unique tasks in ascending content-derived task ID order.
• Canonical training: 2886 tasks, the complete remaining eligible pool after all validation split groups are excluded. There is no separate training target or loaded-record cap.
• Additional orientation variants reserved by validation groups: 201.
• Initial-state/task overlap: 0; orientation-invariant split-group overlap: 0.

Only fairness/evaluation.jsonl is the validation set. candidates.jsonl and reference-solutions.audit-only.json are offline audit material, never rollout observations or runtime validation candidates. The latter contains native action IDs.

Exact semantic restoration

Each row stores initial_state and metadata. The state has exactly these fields: schema, env_id, mission, width, height, max_steps, step_count, grid, agent_pos, agent_dir, and carrying. grid[x][y] is the official MiniGrid [object_type_id, color_id, object_state] encoding in x-major order; empty cells and all walls, goal, key and door are retained. Door states are 0=open, 1=closed/unlocked, 2=locked. carrying is null or the encoded held yellow key. Direction 0/1/2/3 means east/south/west/north. Positions use native (x,y). The mission is exactly use the key to open the door and then get to the goal. Initial step count is zero; max steps is 64. Native actions are deterministic, so generator RNG state is not required to restore a decision or execute a task.

Canonical serialization is UTF-8 JSON with sorted keys, compact separators, ensure_ascii=False, with no added newline for hashing. task_id is SHA256 of the complete initial state. split_group_id hashes that same state omitting only agent_dir, keeping all initial orientation variants in one split. Runtime CPAD matches task ID plus the full current semantic state, including elapsed steps (equivalently remaining budget); actor history is excluded from matching. Object coordinates and solutions are training/audit metadata, never policy text. metadata.image_sha256 hashes raw C-order uint8 RGB pixels of the restored image. No external state resources, seeds-as-task substitutes, or procedural resets are needed.

Physics and observation

Pinned dependencies: MiniGrid 3.0.0, Gymnasium 1.2.3, NumPy 1.26.4. Expose native IDs left=0, right=1, forward=2, pickup=3, toggle=5. Left/right turn. Drop/done are not selectable actions. Invalid responses execute an internal native done/no-op, count one step, and preserve physical state. Supported actions with no observable effect are valid. Native goal reward is 1 - 0.9*t/64 and zero otherwise. Goal success on step 64 wins even if termination and truncation are both true. The environment adds no training penalty.

The official RGBImgObsWrapper(tile_size=56) over the underlying environment with highlight=False produces one full 448×448 RGB image. No recoloring, coordinates, prior images, or reference paths are added. Actor context has mission, direction, holding status, remaining budget and the last two compact action/feedback records. Model/processor weights are not part of this dataset. Experiments use the existing local /data/zhangdw12/models/Qwen2.5-VL-3B-Instruct, with artifact checksums recorded by the experiment; ordinary input/output caps are 2048/256, CPAD input 3072.

Deterministic selection

Selection version minigrid-eval-256-v1. BFS expands the five native MiniGrid primitive actions (including turns, pickup and toggle), with a 64-action limit. Every eligible candidate is restored exactly, checked against its RGB hash, and its shortest path is replayed through the real native environment. Rejections and seed provenance are retained in fairness/provenance.json.

Nearest-rank 25th/50th/75th shortest-length boundaries are [14, 17, 19]. Equal lengths stay together; empty bins are omitted. Length bin is crossed with splitting-wall x and door y. Hamilton largest-remainder allocation assigns 256 slots proportional to stratum size; remainder ties sort by (length_bin, wall_position, door_position). Within strata, ascending (SHA256(UTF8("minigrid-eval-256-v1:" + task_id)), task_id) determines selection. The selected records are sorted by task_id; their whole split groups are removed from training. Audit lists strata, quotas, selected IDs, overlap results and hashes.

Download and verify without the training stack

uvx hf download zhangdw/minigrid --type dataset --revision <published-commit> --local-dir minigrid
cd minigrid
uv venv --python 3.11 .venv
uv pip install --python .venv/bin/python -r requirements.txt
SDL_VIDEODRIVER=dummy .venv/bin/python code/restore_and_verify.py
# Optional: re-run native BFS for every train and validation state.
SDL_VIDEODRIVER=dummy .venv/bin/python code/restore_and_verify.py --bfs

checksums.json verifies the published manifests, audit, provenance, card and restoration sources. No network access is needed after downloading dependencies. Pin the resolved Hub commit, never main, in formal experiment configs.

For complete generation/selection reproduction, use the feat/minigrid-fairness implementation in https://github.com/zhangdw156/verl-agent (issue #20):

python -m examples.data_preprocess.prepare_minigrid_fairness --output generated
python -m examples.data_preprocess.package_minigrid_release --output generated
python -m examples.data_preprocess.prepare_minigrid_fairness --output generated --audit --bfs

The exact generation/state sources are also archived under code/. Running preparation against existing frozen candidates repeats selection byte for byte. A per-seed generation journal permits interrupted generation to resume; it is not part of the published dataset. File hashes identify the exact source bundle.

Manifest SHA256

• train: 9f83a24fb1bdb41987d88294d600175d0e9f9a0742e37c2c432f747740d48b32
• evaluation: 133d3ae56769cb63a783d32e7896db10c4be3c8ccb9acbebd3d05cecb8d5316d
• candidates: e751cac40fba9235baf33535be4d57c6093822e52c36e0999548fabb27f1ef72

MiniGrid is distributed under the MIT license by Farama Foundation. This release contains synthetic task states generated using its DoorKey environment and no personal data. Repository feature acceptance remains pending owner-run ai02 GRPO and CPAD experiments; dataset publication is not a claim of training success.