poset-traces

Branching traces from a polyomino exact-cover solver, used to train POSET — a branching-policy network that orders candidate placements at each branch point.

Source code: https://github.com/CloudHolic/maplestory-union-solver

What's in it

Two related tables, joined by instance_id:

instances config

One row per exact-cover instance. Holds the static structure: the board empty bitmap, the candidate placements (every legal piece position on that board), and a mapping from canonical cell indices to grid indices.

Column	Type	Notes
instance_id	string	unique within the dataset
canonical_bitmaps	list of 36-byte binary	one per piece, canonical orientation
cell_to_grid_idx	list of uint16	canonical → grid mapping
placements	list of struct	each: cells (uint16 list), piece_def_idx (uint16), mark_on_center (bool)

branches config

One row per branch point encountered during the solver's search. Each branch records the pre-state (board state and remaining counts at that node) and the candidate placements the solver considered at that point — which were tried, which succeeded, and the cost of failure (subtree node count for the tried-and-failed ones).

Column	Type	Notes
instance_id	string	foreign key to instances
branch_id	uint32	unique within the instance
pre_state	struct	empty_bitmap (28-byte binary), center_mark (bool), counts (uint8 list)
candidates	list of struct	each: placement_idx (uint32, into instances.placements), tried (bool), succeeded (bool), subtree_nodes (uint64)

A single instance typically yields thousands of branches.

Generation

Generated by a Rust exact-cover solver with cell-MRV backtracking + Luby restart. For each randomly-sampled instance:

1. The solver runs to completion (or timeout, default 10 min).
2. At every branch point, the pre-state and per-candidate outcome are logged. Timed-out instances are dropped.
3. Output is jsonl.gz shards, later converted to this parquet layout by the etl/ pipeline.

Instance distribution is biased toward harder cases (more piece slots, denser board) — the solver is fast on easy instances, so training data benefits from concentrating signal on cases where placement ordering matters most.

Splits

There are no train/test splits in the dataset itself. The training code performs an instance-level hash split (SHA-256 of instance_id, configurable val ratio) at load time. This keeps the dataset reusable for any split decision downstream.

License

MIT.