zhenchonghu/dsaa6000q-a3-curated

--- pretty_name: DSAA6000Q Curated Self-Alignment Dataset tags: - dsaa6000q - instruction-tuning - synthetic - self-alignment --- # DSAA6000Q Curated Self-Alignment Dataset This dataset contains curated instruction-response pairs produced by the standalone notebook pipeline for DSAA6000Q Assignment 3. ## Contents - Rows: 1 - Source file: `/kaggle/working/dsaa6000q_outputs/dsaa6000q_a3/curated/curated_pairs.jsonl` ## Construction 1. Train a backward LoRA model on OpenAssistant Guanaco. 2. Sample 150 single-turn LIMA responses. 3. Generate synthetic instructions with the backward model. 4. Score each pair from 1 to 5 with Qwen3-1.7B. 5. Keep pairs with score >= 4.