saad1926q/qwen3-1.7b-15puzzle-sft-depth-8-15
The saad1926q/qwen3-1.7b-15puzzle-sft-depth-8-15 is a 1.7 billion parameter Qwen3-based model, supervised fine-tuned for 15-puzzle move prediction and reasoning-format alignment. It was trained on optimal 15-puzzle trajectories with per-move rationales for scramble depths 8-15. This model specializes in understanding the 15-puzzle game mechanics and generating structured responses, making it suitable as an SFT initialization for subsequent reinforcement learning in puzzle-solving tasks.
Loading preview...
Overview
This model, saad1926q/qwen3-1.7b-15puzzle-sft-depth-8-15, is a supervised fine-tuned (SFT) checkpoint of the PrimeIntellect/Qwen3-1.7B base model. Its primary purpose is to align the model with the specific format and reasoning required for solving the 15-puzzle game, particularly for move prediction and rationale generation.
Key Capabilities
- 15-Puzzle Move Prediction: Trained to predict optimal moves for 15-puzzle board states.
- Reasoning-Format Alignment: Learns to generate short rationales (
<think>...</think>) before predicting a move (<move>...</move>). - Specific Move Convention: Understands that moves describe the numbered tile moving into the blank, not the blank itself.
- Trajectory Following: Exhibits basic behavior for following optimal 15-puzzle trajectories.
Training Details
The model was fine-tuned on 600 examples from the saad1926q/15-puzzle dataset, focusing on scramble depths 8–15. Each example included a board state, a solver-verified optimal move sequence, and DeepSeek-generated per-move rationales. The training involved full-parameter SFT over 40 steps and 3 epochs, with a sequence length of 3072 and a learning rate of 1e-5.
Intended Use
This checkpoint is primarily intended as an SFT initialization for later Reinforcement Learning (RL) training on the 15-puzzle. It teaches the task format, response structure, legal move semantics, and basic trajectory-following. It is not designed to be a fully capable 15-puzzle solver on its own, especially for deeper scrambles, and may still make suboptimal or illegal moves.