Using LoRA, Full Finetune technique to finetune models in Qwen2.5-Coder series, improved score in several benchmarks (HumanEval, MBPP, DS1000, ...).