xlm.tasks.maze.metrics

Validity-based metrics for maze path generation.

MazeSuccess

Bases: MeanMetric

Measure whether generated paths solve the fixed maze task.

A path succeeds when every decoded cell is traversable, every transition is a four-neighbour move, and all conditioning subgoals occur in order. Inputs are SimpleSpaceTokenizer token IDs by default.

update(pred_cells, subgoal_cells, pred_mask=None)

Add a batch of generated token sequences to the running mean.