pub const fn control_sampling() -> (f32, i32, f32, f32)Expand description
The sampling the control arm applies, on top of a request’s seed.
§Why the temperature alone was not enough
The first version of this control set only CONTROL_TEMPERATURE, and it
failed to degrade anything — measured on Qwen3.5-4B, it scored above
both real arms. The reason is the sampler chain’s order, which ADR 0003
finding 5 measured: llama.cpp applies the truncation samplers before
temperature. With a reasoning recipe’s top_k: 20 and top_p: 0.95
already in force, temperature 2.0 was only flattening a distribution over
twenty surviving tokens — a much tamer change than the number suggests.
So the control disables every truncation sampler as well. A temperature
that cannot be absorbed by a top_k running ahead of it is the only kind
that demonstrates anything.
§It differs from the gglib arm in more than one value, and that is fine
An earlier comment here claimed the control differed in exactly the
temperature, so a gap could only be that. That was already untrue: naming a
temperature claims the coupled trio, so the control’s presence_penalty
and repeat_penalty fall to the class floor rather than matching the
model’s recipe. Isolating one variable is a job for an ablation; this
arm’s job is to be large and known-bad, and breadth serves that.