This model is an online GRPO fine-tuned version of Qwen2.5-Coder-7B-Instruct, trained on the TACO-verified dataset for Python code generation tasks.
Training was performed using LoRA fine-tuning in bfloat16 precision (no quantization) with execution-based reward optimization. During training, multiple candidate programs were generated per prompt, executed against unit tests, and rewarded based on correctness.
After training, the LoRA adapters were merged into the base model using PEFT.
The final checkpoint is fully merged and stored in safetensors format (bf16 precision) and does not require LoRA at inference time.
