| max_iterations: 100 | |
| checkpoint_interval: 10 | |
| log_level: INFO | |
| llm: | |
| primary_model: claude-opus-4-6 | |
| primary_model_weight: 1.0 | |
| api_base: https://api.anthropic.com/v1 | |
| api_key: ${ANTHROPIC_API_KEY} | |
| temperature: 0.7 | |
| max_tokens: 16384 | |
| timeout: 600 | |
| prompt: | |
| system_message: |- | |
| You are an expert in data optimization and LLM prompt caching. | |
| Improve the Evolved class to maximize prefix hit count (PHC) for LLM prompt caching. | |
| Score = 0.95 * avg_hit_rate + 0.05 * speed_score. | |
| Focus on optimal column ordering per row, prefix-aware greedy approaches, vectorized operations. | |
| pandas, numpy are available. Keep the Evolved class structure and reorder method signature. | |
| num_top_programs: 3 | |
| use_template_stochasticity: true | |
| database: | |
| population_size: 50 | |
| archive_size: 25 | |
| num_islands: 3 | |
| elite_selection_ratio: 0.3 | |
| exploitation_ratio: 0.7 | |
| evaluator: | |
| timeout: 360 | |
| cascade_evaluation: false | |
| parallel_evaluations: 1 | |
| use_llm_feedback: false | |
| diff_based_evolution: true | |
| allow_full_rewrites: false | |