code/self_healing_finetuning/training/iteration_policy.py

from __future__ import annotations


DEFAULT_ITERATION_POLICY = {
    "max_iterations_per_run": 5,
    "max_auto_repair_iterations": 3,
    "max_total_iterations_per_candidate": 8,
    "stop_on_first_certified_candidate": True,
    "require_human_approval_after_auto_repairs": True,
    "require_new_data_or_new_repair_reason_after_failed_loop": True,
}


def should_stop(iteration: int, failures: int, certified: bool) -> tuple[bool, str]:
    if certified and DEFAULT_ITERATION_POLICY["stop_on_first_certified_candidate"]:
        return True, "candidate_ready_for_paired_eval"
    if iteration >= DEFAULT_ITERATION_POLICY["max_iterations_per_run"]:
        return True, "max_iterations_reached"
    if failures >= 2:
        return True, "repeated_same_failure_count"
    return False, "continue"