Sokoban fairness: 5,000 training boards and 256 fixed validation boards

Frozen, solvable 6×6, one-box Sokoban boards for comparing visual-agent training methods such as GRPO and CPAD. The environment renders each stored board into its ordinary RGB game image. Models interact with those images and choose up, down, left, or right. These JSONL files contain environment states, not image/action demonstrations or the trainer's placeholder Parquet.

File	Tasks	Purpose
fairness/train.jsonl	5,000	Fixed training pool, unchanged
fairness/evaluation.jsonl	256	Canonical fixed validation set
fairness/evaluation-256.selection.json	—	Selection rule, quotas, task IDs and checksums
fairness/summary.json	—	Current split counts and statistics
fairness/verification.json	—	RGB replay and shortest-path verification of all 256 boards
fairness/evaluation-pool-1000.jsonl	1,000	Original source pool, retained for reproducibility
fairness/source-pool-summary.json	—	Original generation audit
fairness/source-pool-verification.json	—	Original verification of 5,000 + 1,000 boards

The default fairness configuration uses 256 validation boards. source_pool_1000 is an archive configuration, not the current evaluation protocol.

Fixed proportional selection

The evaluation protocol sokoban-eval-256-v1 selects a subset of the original 1,000-board pool at revision e84324719b3587178b40c8597dfc892c2d67172d. The selection uses shortest solution length in player actions, including walking, measured by exact shortest-path search.

Shortest solution	Source boards	Source proportion	Selected boards
1–4 actions	531	53.1%	136
5–8 actions	412	41.2%	105
9–14 actions	57	5.7%	15
Total	1,000	100%	256

Quotas use Hamilton's largest-remainder rule. Within each stratum, rank by SHA256(UTF8("sokoban-eval-256-v1:" + task_id)) ascending, breaking hash ties by task ID. Select the first quota-sized subset, then write all selected rows in their original source order. Each selected line is byte-identical to its source line. The rule does not depend on model outputs, success rates, training seeds or method identity. The selected IDs are frozen; no resampling occurs between methods, checkpoints or runs. The reproducible selector is fairness/select_evaluation_256.py.

Shortest solution length is an objective board feature, not a measurement of VL-model difficulty. This is a custom proportional split, not an official Sokoban benchmark or another paper's evaluation protocol. Selection does not force representation of every exact path length; the selected lengths span 1–11 actions, while the source pool spans 1–14.

Boards are unique within and across training and validation, including under rotations and reflections of the complete initial state (walls, targets, boxes and player). This does not exclude shared wall layouts or all states reachable later during play. Generation uses the existing 6×6/one-box defaults, search depth 30 and 20 topology-generation steps. Training candidates start at seed 0, source-pool candidates at 65,536. All boards are nonterminal and solvable.

Runtime protocol

Use env.env_name=Sokoban, env.fairness=true, env.sokoban.mode=rgb_array, env.sokoban.dim_room=[6,6], env.sokoban.num_boxes=1, and env.sokoban.search_depth=30 with a compatible VL model.

Runtime resets restore the complete stored board, independent of generator changes. Training samples without replacement within each update using a private environment RNG. Each selected task is repeated env.rollout.n times from the same board. Validation runs all 256 fixed boards once each, in stored order. With 128 workers, each validation is two chunks of 128. The placeholder Parquet stays at 128 rows with data.val_batch_size=128 and actor_rollout_ref.rollout.val_kwargs.n=1; its size does not limit coverage.

Validation and training workers are phase-exclusive. Validation does not advance the training sampler. The sampler survives in-process worker releases; checkpoint resumption does not currently restore its state.

Match initial VL weights, image processing, action horizon, reward, rollout group size, update budget and decoding settings across methods. The current GRPO/CPAD launchers use a 15-action horizon and evaluate before training and every five updates. CPAD's additional scoring cost should be reported separately.

Each row contains a shortest solution for audit only. Runtime observations and CPAD teacher prompts never include the stored solution, generation seed or shortest-path length. CPAD derives preferences from training rollouts.

Runtime update required: fairness/evaluation.jsonl now contains 256 rows. The verl-agent loader must expect 256 and SHA-256 8b249dff58cd53e97769b8ceac418ea458a92fd93d4f50f702c018b5aff41fad. An older loader expecting 1,000 rows, or an old cached evaluation file, will fail checksum validation. Update the code and download the current evaluation file into the directory used by the experiment. Training data is unchanged.

See fairness/README.md for schema, download and reproduction.