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.
Keep examples with score >= 4.
train.jsonl: curated high-quality examples for final instruction tuning
scores.jsonl: all 150 scored… See the full description on the dataset page:
https://huggingface.co/datasets/sunming-giegie/assignment3-lima-curated-150.
