This dataset contains the curated instruction-response pairs produced in Part 3 of Assignment 3.
Start from LIMA single-turn examples.
Use the backward model to infer instructions from responses.
Score each (generated_instruction, response) pair with Qwen/Qwen3-1.7B using few-shot prompting and a 1-5 quality rubric.
Keep examples with score >= 4.
train.jsonl: curated high-quality examples for final… See the full description on the dataset page:
https://huggingface.co/datasets/sunming-giegie/assignment3-curated-lima-dataset.
