Tuner

chimera.training.tuner provides hyperparameter search for synthesis configurations. SynthesisTuner performs grid search over a SearchSpace, running synthesis for each configuration and selecting the best one.

---

Key Classes

SearchSpace

Defines the hyperparameter search space with categorical choices. Calling configurations() produces the full Cartesian product (grid search).

space = SearchSpace()
space.choice("model", ["glm-5", "claude-sonnet-4-20250514"])
space.choice("max_steps", [10, 25])
# Produces 4 configurations

SynthesisTuner

Grid search over synthesis configurations. Creates a fresh environment for each trial, runs synthesis, and picks the best result.

Constructor parameters:

Param	Type	Description
spec	Spec	The synthesis specification
env_factory	Callable[[], Environment]	Returns a fresh env for each trial
agent_factory	Callable[[dict], Agent] | None	Optional custom agent builder

Key method: search(space, max_trials, metric, callbacks) -> TunerResult.

TunerResult

Field	Type	Description
best_config	dict	Configuration that achieved the highest score
best_score	float	Score of the best configuration
trials	list[TrialResult]	All trial results in execution order
total_cost	float	Sum of costs across all trials

Usage

from chimera.training.tuner import SynthesisTuner, SearchSpace

space = SearchSpace()
space.choice("model", ["glm-5", "claude-sonnet-4-20250514"])
space.choice("max_iterations", [5, 10, 20])

tuner = SynthesisTuner(
    spec=spec,
    env_factory=lambda: LocalEnvironment(workdir="/tmp/trial", test_cmd="pytest"),
)

result = tuner.search(space, max_trials=6, metric="pass_rate")
print(f"Best config: {result.best_config}")
print(f"Best score:  {result.best_score:.0%}")
print(f"Total cost:  ${result.total_cost:.2f}")

Related

• Training concepts — strategies and the Trainer
• Cost Tracking — per-model pricing