saad1926q/8-puzzle

8-puzzle / 3x3 sliding puzzle

Fixed 3x3 / 8-puzzle evaluation data and teacher-rollout datasets for sliding-puzzle reasoning experiments.

Configs

• —eval: 31 fixed evaluation puzzles, with exactly one puzzle at every optimal distance from 1 through 31.
• —sft-source: 200 fresh boards, balanced with 20 boards at each exact depth 10-19. It excludes the reserved evaluation boards and deliberately contains no solution-action labels.
• —sft: 3,745 replay-verified, fully annotated teacher decisions from solved sft-source trajectories.

The eval configuration uses the eval split. sft-source and sft use the train split.

Fields

• —board: flat 3x3 board, with 0 as the blank tile.
• —scramble_depth: exact optimal solution length for the generated evaluation set.
• —bucket: easy, medium, or hard.
• —action_interface: schema identifier, tile_id_v1.
• —optimal_actions: integer tile IDs to slide from the board to the solved state.
• —optimal_length: length of optimal_actions.

Evaluation set

The evaluation set contains one puzzle at every exact optimal distance from 1 through 31. The boards are sampled reproducibly from the reachable state space and are kept separate from the local SFT datasets. Every stored action sequence is replay-validated and optimal.

Actions are numbered tiles from 1 through 8; each action slides that tile into the blank.

sft-source fields

• —id: stable source-board identifier.
• —board: flat 3x3 board, with 0 as the blank tile.
• —optimal_length: exact optimal solution length; source metadata only and not a teacher-model input.
• —action_interface: schema identifier, tile_id_v1.

sft-source intentionally omits optimal_actions; teacher models must produce their own trajectories.

sft fields

• —prompt: evaluator-format chat messages containing the system prompt, current board, and up to four prior rationale/action/result turns.
• —completion: one assistant message containing the annotated rationale and one verified slide_tile tool call.
• —tools: the slide_tile function schema used by the evaluator.
• —metadata: replay and audit fields including source_id, target_turn, board, tile, next_board, and history_turns.

The SFT training records were generated with a fixed history window of four turns.