gokul8967/snake-sft-8x8

Snake SFT Dataset

A dataset of 2,000 Snake gameplay episodes with step-by-step reasoning and planning.

Dataset Statistics

• —Episodes: 2,000
• —Training examples: 48,429
• —Board size: 8x8
• —Average apples per episode: 24.2
• —Episodes with 25+ apples: 984 (49.2%)

Data Format

Each line contains a JSON object with:

• —task: Game identifier
• —answer: Episode summary (apples eaten, death status, steps)
• —reward: Number of apples eaten
• —prompt: System prompt for the AI
• —completion: Alternating user/assistant messages showing the gameplay

User Messages

Contain <APPLE id=X> observations with:

• —Snake head position
• —Snake body coordinates
• —Apple location
• —Board size

Assistant Messages

Contain <THINK> reasoning blocks with:

• —Path coordinates to the apple
• —Move sequence
• —Tail safety calculations
• —Safety verification

Followed by <ACTIONS> with the actual moves to execute.

Usage

This dataset is designed for fine-tuning language models to play Snake with sophisticated planning and reasoning capabilities.

License

Apache 2.0