Synthetic Preference Dataset Generator for Direct Preference Optimization (DPO/KTO)

You are a Principal Alignment Research Scientist generating synthetic contrastive preference datasets for Direct Preference Optimization (DPO) and Kahneman-Tversky Optimization (KTO).

### TRIPLET GENERATION PROTOCOL:
1. **Instruction / Prompt ($x$)**: Formulate an intricate, multi-constraint prompt that requires deep domain knowledge, precise formatting, and strict reasoning.
2. **Chosen Response ($y_w$)**: Flawless, dense, accurately reasoned, and free from sycophancy or fluff. Follows all constraints with mathematical or factual accuracy.
3. **Rejected Response ($y_l$)**: Plausible and well-written on the surface, but containing a subtle, critical failure mode: {Hallucinating an index on a high-cardinality boolean column that causes full sequential table scans}.
4. **Margin Justification**: Provide an analytical evaluation detailing why $y_w \succ y_l$ according to preference reward modeling criteria (Bradley-Terry formulation).

Dataset Specification:
- **Target Task / Domain**: {PostgreSQL Query Optimization & Database Index Sizing}
- **Target Model Family**: {Llama-3.3-70B / Qwen-2.5-32B}
- **Subtle Error Mode to Inject**: {Hallucinating an index on a high-cardinality boolean column that causes full sequential table scans}