pub struct TuneConfig {
pub model_id: i64,
pub task_suite: TaskSuite,
pub sweep: SweepSpec,
pub seed_from_gguf: bool,
pub seed_from_family_presets: bool,
pub weights: ScoreWeights,
pub prune_fraction: f32,
pub ctx_size: Option<u64>,
}Expand description
Configuration for a tune benchmark run.
A tune run evaluates one model against many candidate InferenceConfig
sampling settings, scoring each candidate against an agentic tool-calling
task suite to find the settings that make the model both accurate at
tool calls and resistant to loop/stagnation guard triggers.
Fields§
§model_id: i64Database ID of the model to tune.
task_suite: TaskSuiteTask suite to evaluate candidates against.
sweep: SweepSpecSampling-parameter values to sweep (cartesian product forms the grid).
seed_from_gguf: boolSeed additional candidates from the model’s GGUF metadata author-recommended sampling defaults, when present.
seed_from_family_presets: boolSeed additional candidates from the built-in per-model-family preset table (e.g. Qwen coding-mode defaults).
weights: ScoreWeightsWeights used to combine per-candidate metrics into a composite score.
prune_fraction: f32Fraction of candidates dropped after the cheap pre-screen round.
0.5 drops the bottom half of candidates before running the full
task suite on the remaining survivors. Clamped to [0.0, 0.9].
ctx_size: Option<u64>Override the llama-server context window size for this run.
Implementations§
Source§impl TuneConfig
impl TuneConfig
const fn default_prune_fraction() -> f32
Trait Implementations§
Source§impl Clone for TuneConfig
impl Clone for TuneConfig
Source§fn clone(&self) -> TuneConfig
fn clone(&self) -> TuneConfig
1.0.0 (const: unstable) · Source§fn clone_from(&mut self, source: &Self)
fn clone_from(&mut self, source: &Self)
source. Read more