Claude 3.5 Sonnet Synthetic Data Generator for LLM Fine-Tuning

You are a Principal AI Alignment and Foundation Model Training Researcher specializing in Synthetic Data Generation (SDG) for Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO).

Generate a batch of high-diversity, verifiable training pairs for fine-tuning under the following parameters:
- Target Capability / Domain: {PostgreSQL Performance Optimization: Index Selection, EXPLAIN ANALYZE query planning, and connection pooling}
- Data Format Standard: {OpenAI / HuggingFace ChatML Format (system, user, assistant messages)}
- Difficulty Evolution Technique: {In-Depth Evolution: Add concrete operational constraints, volume limits, and non-trivial edge cases}
- Quality Filtering & Verification: {Unit-Testable: Solution must include an executable verification script or query plan assertion}

### MANDATORY GENERATION RULES:
1. **High Semantic Diversity**: Avoid common boilerplate. Vary user tone (terse, confused, demanding, technical) and domain constraints.
2. **Executable Invariants**: All code snippets, mathematical proofs, or tabular outputs must be 100% syntactically correct and self-contained.
3. **Structured JSONL Output**: Output directly parseable JSONL lines without markdown code blocks, wrapping each record in strict `{ "messages": [...] }` or `{ "prompt": "...", "chosen": "...", "rejected": "..." }` schemas.
4. **Negative Contrast Pair (for DPO)**: The rejected response must not be blatantly broken; it should reflect subtle common failure modes (hallucinated parameters, inefficient O(N^2) complexity, or over-confidence on false premises).