Regularization

chimera.training.regularization provides a callback that combines test pass-rate with critic-based code quality scoring. This lets strategies prefer simpler, higher-quality solutions among those that pass the tests.

---

Key Classes

RegularizationCallback

A Callback that scores code quality after each epoch using a Critic. When the epoch reaches a minimum pass rate, the critic evaluates the generated code on readability, maintainability, and simplicity.

Constructor parameters:

Param	Type	Default	Description
critic	Critic	required	Critic used to evaluate code quality
weight	float	0.3	Weight given to the critic score (0.0-1.0)
min_pass_rate	float	0.5	Minimum pass rate before critic runs

Key method:

• combined_score(pass_rate, critic_score) -> float — returns pass_rate * (1 - weight) + critic_score * weight.

Usage

from chimera.training.regularization import RegularizationCallback
from chimera.critic.llm_critic import LLMCritic

critic = LLMCritic(provider=provider)
reg = RegularizationCallback(critic=critic, weight=0.3, min_pass_rate=0.5)

# Use as a callback during synthesis
result = trainer.synthesize(
    strategy=TestConvergence(max_iterations=20),
    callbacks=[reg],
)

# Compute a blended score manually
score = reg.combined_score(pass_rate=0.95, critic_score=0.8)
# => 0.95 * 0.7 + 0.8 * 0.3 = 0.905

Related

• Critic module — LLMCritic and ChecklistCritic
• Training concepts — callbacks and strategies