ion7-llm / sampler.profiles

module

sampler.profiles

Functions

M.balanced

General-purpose default. Mild top-k / top-p / min-p with a moderate temperature ; a reasonable choice when you do not have a specific reason to deviate.

M.balanced()
→ ion7.core.Sampler

M.precise

Tighter cut-offs, lower temperature. Best when you want plausible, on-topic continuations without much variability.

M.precise()
→ ion7.core.Sampler

M.creative

Wider candidate window, higher temperature. For brainstorming, creative writing, idea generation.

M.creative()
→ ion7.core.Sampler

M.code

Code-tuned. Mild repetition penalty, low temperature, narrow top-k. Reduces the rate of accidental loops in code completions.

M.code()
→ ion7.core.Sampler

M.fast

Pure greedy. Deterministic and the fastest possible step ; best for grammar-constrained workloads where any randomness would conflict with the constraint.

M.fast()
→ ion7.core.Sampler

M.thinking

Tuned for reasoning models — wider top-p, slightly cooler temp, still some entropy. Pair with `engine.chat({ thinking = true })`.

M.thinking()
→ ion7.core.Sampler