Datasets:

albertoRodriguez97
/

MazeBench

@@ -12,29 +12,20 @@ tags:
   - vision-language
   - benchmark
 size_categories:
-  - 10K<n<100K
 ---
 # MazeBench
-A large-scale visual maze reasoning dataset for evaluating and training multimodal language models.
-**Paper:** [From Pixels to BFS: High Maze Accuracy Does Not Imply Visual Planning](https://arxiv.org/abs/2603.26839)
 **Code:** [github.com/alrod97/LLMs_mazes](https://github.com/alrod97/LLMs_mazes)
 ## Overview
-MazeBench contains ~65,000 procedurally generated maze images spanning 8 structural families, 5 grid sizes (5x5 to 20x20), and balanced reachable/unreachable splits. Each record includes a maze PNG, a chat-formatted prompt, and a ground-truth JSON answer with the shortest path.
-## Splits
-| Split | Records | Purpose |
-| --- | ---: | --- |
-| `train` | ~50,000 | Training |
-| `val_iid` | ~2,000 | In-distribution validation (same families and palettes as train) |
-| `val_ood` | ~4,000 | Out-of-distribution validation (held-out palettes and families) |
-| `test` | ~9,000 | Held-out evaluation |
 ## Maze Families
@@ -47,62 +38,16 @@ MazeBench contains ~65,000 procedurally generated maze images spanning 8 structu
 - **near_goal_blocked** — goal is close but path is indirect
 - **near_start_blocked** — start area has limited exits
-## Record Format
-Each JSONL line contains:
-```json
-{
-  "id": "train_00001",
-  "image_path": "train/images/train_00001.png",
-  "messages": [
-    {"role": "system", "content": "..."},
-    {"role": "user", "content": "..."},
-    {"role": "assistant", "content": "{\"reachable\": true, \"shortest_path_length\": 12, \"path\": [\"R\",\"D\",...]}"}
-  ],
-  "meta": {
-    "family": "bottleneck",
-    "grid_size": [10, 10],
-    "reachable": true,
-    "shortest_path_length": 12,
-    "palette": "forest"
-  }
-}
-```
-## Directory Structure
-```
-train/
-  train.jsonl
-  images/
-    train_00001.png
-    ...
-val_iid/
-  val_iid.jsonl
-  images/
-val_ood/
-  val_ood.jsonl
-  images/
-test/
-  test.jsonl
-  images/
-metadata.json
-```
 ## Usage
-```python
-from datasets import load_dataset
-ds = load_dataset("albertoRodriguez97/MazeBench")
-```
-Or download directly:
 ```bash
 pip install huggingface_hub
-huggingface-cli download albertoRodriguez97/MazeBench --repo-type dataset --local-dir dataset/
 ```
 ## Citation
@@ -115,7 +60,3 @@ huggingface-cli download albertoRodriguez97/MazeBench --repo-type dataset --loca
   year    = {2026},
 }
 ```
-## License
-MIT

   - vision-language
   - benchmark
 size_categories:
+  - n<1K
 ---
 # MazeBench
+The evaluation set from the paper *From Pixels to BFS: High Maze Accuracy Does Not Imply Visual Planning*.
+**Paper:** [arXiv:2603.26839](https://arxiv.org/abs/2603.26839)
 **Code:** [github.com/alrod97/LLMs_mazes](https://github.com/alrod97/LLMs_mazes)
 ## Overview
+110 procedurally generated maze images spanning 8 structural families and grid sizes from 5x5 to 20x20, with ground-truth shortest-path annotations. These are the mazes used to evaluate multimodal LLMs in the paper.
 ## Maze Families
 - **near_goal_blocked** — goal is close but path is indirect
 - **near_start_blocked** — start area has limited exits
+## Contents
+- `gen_maze_001.png` ... `gen_maze_110.png` — 1024x1024 maze images
+- `maze_annotations.json` — ground-truth annotations (reachability, shortest path length, accepted shortest paths)
 ## Usage
 ```bash
 pip install huggingface_hub
+huggingface-cli download albertoRodriguez97/MazeBench --repo-type dataset --local-dir generated_mazes/
 ```
 ## Citation
   year    = {2026},
 }
 ```