qwen3-1.7b-15puzzle-sft-depth-8-15/README.md

license, base_model, tags, datasets

license	base_model	tags	datasets
apache-2.0	PrimeIntellect/Qwen3-1.7B	qwen3 15-puzzle sft reasoning puzzle-solving	saad1926q/15-puzzle

Qwen3 1.7B 15-Puzzle SFT

This is a supervised fine-tuned checkpoint of PrimeIntellect/Qwen3-1.7B for 15-puzzle move prediction and reasoning-format alignment.

The model was fine-tuned on solved 15-puzzle trajectories from saad1926q/15-puzzle, using depth 8–15 puzzles.

Each training example contains:

• A board state
• A solver-verified optimal move sequence
• Short per-move rationales

Training Data

Dataset split

• Dataset: saad1926q/15-puzzle
• Config/subset: sft
• Split: sft
• Rows: 600
• Scramble depths: 8–15
• Base trajectories: Solver-verified optimal paths
• Rationale source: DeepSeek-generated rationales with tile-move convention

Move Convention

Moves describe the numbered tile moving into the blank, not the blank moving.

For example, if the board is:

13 _ 14 15

then

<move>left</move>

means tile 14 moves left into the blank, producing:

13 14 _ 15

Intended Use

This checkpoint is primarily intended as an SFT initialization for later RL training on the 15-puzzle.

The goal of SFT is to teach:

• Task format
• The <think>...</think> and <move>...</move> response structure
• Legal move semantics
• Basic trajectory-following behavior

It should not be treated as a fully capable 15-puzzle solver, especially on deeper scrambles.

Response Format

The model is trained to respond with:

<think>
short rationale
</think>
<move>left</move>

Allowed moves are:

• up
• down
• left
• right

Training

Base model

• PrimeIntellect/Qwen3-1.7B

Fine-tuning framework

• PrimeIntellect Prime-RL SFT

Training setup

• Full-parameter SFT
• 600 examples
• 40 training steps
• 3 epochs
• Sequence length: 3072
• Learning rate: 1e-5
• Final loss: approximately 0.47

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "saad1926q/qwen3-1.7b-15puzzle-sft-depth-8-15"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
)

Example Prompt

You are Player 0 in 15-puzzle.

A 4x4 sliding puzzle board is given. The blank tile is shown as _.

Your goal is to reach the solved board:

1 2 3 4
5 6 7 8
9 10 11 12
13 14 15 _

At each turn, choose one legal move.

Moves describe the numbered tile moving into the blank, not the blank moving.

Allowed moves are:
up, down, left, right.

Wrap your move in <move>...</move>.

Current board:

1 2 3 4
5 6 7 8
9 10 11 _
13 14 15 12

Expected style:

<think>
Move tile 12 up into the blank to place it in its correct final position.
</think>
<move>up</move>

Limitations

• This is an SFT checkpoint, not an RL-trained solver.
• The model may still make illegal or suboptimal moves.
• The training set is small and focused on scramble depths 8–15.
• The rationales are intended to teach format and move semantics, not guarantee perfect search behavior.