stablegradients/maze-17x17-500k-general

---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- maze
- reasoning
- reinforcement-learning
- sft
size_categories:
- 100K<n<1M
configs:
- config_name: binary
data_files:
- split: train
path: train_binary.parquet
- split: test
path: test_binary.parquet
- config_name: continuous
data_files:
- split: train
path: train_continuous.parquet
- split: test
path: test_continuous.parquet
- config_name: distance
data_files:
- split: train
path: train_distance.parquet
- split: test
path: test_distance.parquet
---
# Maze 17x17 500k — General Paths

Supervised fine-tuning corpus of 17x17 mazes with the target being a valid (not necessarily shortest) path from START to GOAL. Mazes are generated with Prim's algorithm.

## Splits and configs

- **train**: 450,000 examples
- **test**: 512 examples

Three reward configurations are provided — they share the same prompts but
differ in the reward-model metadata used by downstream RL:

| config | reward signal |
|--------------|-------------------------------------------|
| `binary` | 1 if the path reaches the goal, else 0 |
| `continuous` | fractional progress toward the goal |
| `distance` | goal reached + solution-quality component |

## Schema

Each row has columns: `data_source`, `prompt`, `ability`, `reward_model`,
`extra_info`. `prompt` is a chat-formatted list of messages whose user turn
contains the maze grid in a tokenized form (`WALL`, `PATH`, `START`, `GOAL`,
`NEWLINE`). The target trajectory lives in `extra_info.answer` and
`reward_model.ground_truth`.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("stablegradients/maze-17x17-500k-general", "binary")
print(ds["train"][0])
```