| license: mit | |
| task_categories: | |
| - other | |
| tags: | |
| - maze | |
| - world-model | |
| - sequence-modeling | |
| - transformer | |
| - mamba | |
| pretty_name: World Model For Maze | |
| # WorldModelForMaze | |
| Code, datasets, and trained checkpoints for studying world-model representations in maze navigation, based on a modified NanoGPT. | |
| ## Contents | |
| - `*.py` — training, testing, probing, and visualization scripts (see `readme.md`). | |
| - `model/` — architectures: transformer, transformer-rope, transformer-nextlat, mamba, mamba2, gated-deltanet, gru. | |
| - `data/maze/100/` — tokenized maze datasets for Tasks A/C/E/H/I (RWs paths, 100 nodes). | |
| - `out/` — final (10000-iter) checkpoints, `maze_kdetour` results, and plots. | |
| > Intermediate (non-10000-iter) checkpoints (`out2/`) are not included here. | |
| ## Usage | |
| See `readme.md` for full instructions on data generation, training, testing, probing, and the k-step detour analysis. | |
| Quick start: | |
| ```bash | |
| git clone https://huggingface.co/datasets/Kalso42/WorldModelForMaze | |
| cd WorldModelForMaze | |
| python train_maze.py --tasks C1 --path_type RWs --num_train_dataset 10M --config 6_6_384 | |
| ``` | |