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

---
license: apache-2.0
base_model: PrimeIntellect/Qwen3-1.7B
tags:
  - qwen3
  - 15-puzzle
  - sft
  - reasoning
  - puzzle-solving
datasets:
  - 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`](https://huggingface.co/datasets/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:

```text
13 _ 14 15
```

then

```xml
<move>left</move>
```

means tile **14** moves left into the blank, producing:

```text
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:

```xml
<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

```python
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

```text
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:

```xml
<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.