WorldModelForMazeWithX

Maze pathfinding sequences for training/probing sequence models (Transformer, Mamba, GRU, Gated-DeltaNet, ...). Task C1: relative-turn navigation on a fixed 10×10 directed grid. Includes a special x terminator marking wall-hit (illegal) paths, used to study a model's ability to recognize its own errors.

Maze

• 10×10 grid, 100 nodes (0–99). Directed edges (down/right, both directions added), edge probability 0.6.
• Graph: maze_graph_C1_RWs.graphml (node ids are strings '0'..'99').

Sequence format (Task C1)

C <src> <tgt> : <turn tokens...>

• Agent starts facing East. Each turn token both rotates and advances one cell:
  • F forward, L left, R right, T turn-around.
• Nodes 0–99 are single tokens; : separates the prompt from the path.
• x: wall-hit terminator. A "wrong" path is a correct path with one turn corrupted into a wall direction, then x appended. Only the final x is supervised during training.

Files (under data/)

file	rows	wrong (x) ratio	note
train_C1_RWs_5M.txt / .bin	5,000,000	0.0	all-correct paths
train_C1_RWs_8M.txt / .bin	8,000,000	0.0	all-correct paths
train_C1_RWs_10M.txt / .bin	12,000,000	0.2	2M wall-hit paths ending in x
test_C1_RWs_10K.txt	~10K	—	held-out test prompts
val_C1_RWs_10K.bin	—	—	tokenized validation
meta_C1_RWs.pkl	—	—	vocab/stoi/itos (vocab size 132, x id 131)
maze_graph_C1_RWs.graphml	—	—	the maze graph

• .txt: human-readable sequences. .bin: uint16 token stream (read with numpy.memmap). .bin can be regenerated from .txt via prepare_multitask_minigpt.py.
• Only the 10M set contains wrong paths; 5M / 8M never expose x (models trained on them never emit x).

Loading

import pickle, numpy as np
meta = pickle.load(open("data/meta_C1_RWs.pkl", "rb"))
itos = meta["itos"]                       # x token id = 131
ids = np.memmap("data/train_C1_RWs_10M.bin", dtype=np.uint16, mode="r")
print(" ".join(itos[i] for i in ids[:60]))

Citation

Generated with data/maze/create_multitask_maze.py (--tasks C1 --path_type RWs, --wrong_ratio 0.2 for 10M, 0.0 for 5M/8M).