Qwen3-4B-Unique-Solutions
This model is a fine-tuned version of Qwen3-4B. It was trained using Reinforcement Learning (GRPO) to generate structurally unique code solutions instead of repeating the same template.
Training details
- Trained for 1000 steps on the mbpp dataset.
- Hardware: Consumer GPU (RTX 5070 Laptop) with 8 GB VRAM using 4-bit quantization.
- Method: A penalty was applied for duplicate code structures, and a bonus was given for new, correct algorithms.
Links
- Full code, evaluator, and training scripts: GitHub Repository
- Downloads last month
- 24
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support