archit11/

track_b_sft

A synthetic instruction-tuning dataset for Python coding tasks, generated from the verl open-source RL framework codebase. Used to fine-tune Qwen/Qwen2.5-Coder-1.5B with LoRA, achieving +24 point absolute improvement in pass@1 (0.565 → 0.804) after 3 epochs of training on a T4 GPU.

	Dataset Summary




	

Split Examples

train 512

test 94

total 303

	Task Categories




	

Category Count Description… See the full description on the dataset page: https://huggingface.co/datasets/archit11/track_b_sft.