gglib_core/domain/benchmark/
mod.rs1#![doc = include_str!("README.md")]
2
3pub mod agentic;
4pub(crate) mod compare;
5pub(crate) mod events;
6pub(crate) mod perf;
7pub mod run;
8pub(crate) mod summary;
9pub mod tune;
10
11pub use agentic::{
12 AgenticEvalConfig, AgenticEvalReport, AgenticTaskComparison, ArmDelta, ArmScores,
13 CONTROL_MIN_COMPOSITE_GAP, CONTROL_MIN_P, CONTROL_TEMPERATURE, CONTROL_TOP_K, CONTROL_TOP_P,
14 ControlVerdict, DEFAULT_SEEDS, DeltaWithheld, EFFECT_NOISE_RATIO, EffectVerdict, EvalArm,
15 REPLICATE_SEED_OFFSET, control_sampling, replicate_seeds,
16};
17pub use compare::{CompareConfig, ModelCompareResult};
18pub use events::{BenchmarkEvent, BenchmarkModelResult};
19pub use perf::{ModelPerfResult, PerfConfig};
20pub use run::{BenchmarkRun, BenchmarkRunStatus, BenchmarkRunType};
21pub use summary::ModelBenchmarkSummary;
22pub use tune::{
23 CandidateSource, GeneratedOutput, ScoreWeights, SweepSpec, TaskCategory, TaskSuite,
24 TuneCandidateResult, TuneConfig, TuneTask, TuneTaskResult,
25};