{"id":"llm-tuning-patterns","name":"llm-tuning-patterns","summary":"LLMチューニングパターン","body":"# LLM Tuning Patterns\n\nEvidence-based patterns for configuring LLM parameters, based on APOLLO and Godel-Prover research.\n\n## Pattern\n\nDifferent tasks require different LLM configurations. Use these evidence-based settings.\n\n## Theorem Proving / Formal Reasoning\n\nBased on APOLLO parity analysis:\n\n| Parameter | Value | Rationale |\n|-----------|-------|-----------|\n| max_tokens | 4096 | Proofs need space for chain-of-thought |\n| temperature | 0.6 | Higher creativity for tactic exploration |\n| top_p | 0.95 | Allow diverse proof paths |\n\n### Proof Plan Prompt\n\nAlways request a proof plan before tactics:\n\n```\nGiven the theorem to prove:\n[theorem statement]\n\nFirst, write a high-level proof plan explaining your approach.\nThen, suggest Lean 4 tactics to implement each step.\n```\n\nThe proof plan (chain-of-thought) significantly improves tactic quality.\n\n### Parallel Sampling\n\nFor hard proofs, use parallel sampling:\n- Generate N=8-32 candidate proof attempts\n- Use best-of-N selection\n- Each sample at temperature 0.6-0.8\n\n## Code Generation\n\n| Parameter | Value | Rationale |\n|-----------|-------|-----------|\n| max_tokens | 2048 | Sufficient for most functions |\n| temperature | 0.2-0.4 | Prefer deterministic output |\n\n## Creative / Exploration Tasks\n\n| Parameter | Value | Rationale |\n|-----------|-------|-----------|\n| max_tokens | 4096 | Space for exploration |\n| temperature | 0.8-1.0 | Maximum creativity |\n\n## Anti-Patterns\n\n- **Too low tokens for proofs**: 512 tokens truncates chain-of-thought\n- **Too low temperature for proofs**: 0.2 misses creative tactic paths\n- **No proof plan**: Jumping to tactics without planning reduces success rate\n\n## Source Sessions\n\n- This session: APOLLO parity - increased max_tokens 512->4096, temp 0.2->0.6\n- This session: Added proof plan prompt for chain-of-thought before tactics","author":"@parcadei","ownerProfile":null,"authorContacts":null,"sourceUrl":"https://github.com/parcadei/Continuous-Claude-v3/tree/main/.claude/skills/llm-tuning-patterns","license":"MIT","category":"writing","lang":"en","tokens":437,"stars":0,"calls30d":1,"claimed":false,"visibility":"public","origin":"crawler","version":"0.1.0","createdAt":"2026-08-22","updatedAt":"2026-08-22","files":[],"requires":{"mcp":[],"tools":[]},"safety":{"flags":[],"scannedAt":"2026-08-22","hasScripts":false,"networkEndpoints":[]}}