Настройка параметров LLM по типу задачи

ai-tooling

llm-tuning-patterns is a Claude Code skill that provides evidence-based LLM parameter configurations for different task types, derived from APOLLO and Godel-Prover research. For theorem proving in Lean 4, it specifies temperature 0.6, top_p 0.95, and max_tokens 4096 to support chain-of-thought reasoning and proof planning before tactics; code generation calls for temperature 0.2–0.4 and 2048 tokens; creative tasks warrant temperature 0.8–1.0. The skill also covers parallel sampling — generating N=8–32 candidate proofs at temperature 0.6–0.8 and applying best-of-N selection for hard theorems. Documented anti-patterns include truncating chain-of-thought with a 512-token limit and using low temperature for formal reasoning, which narrows the space of viable proof tactics. Useful for developers and researchers embedding LLMs into formal verification or engineering workflows.