Skip to content

FullEvaluationPolicy

gepa.strategies.eval_policy.FullEvaluationPolicy

Bases: EvaluationPolicy[DataId, DataInst]

Policy that evaluates all validation instances every time.

Functions

get_eval_batch(loader: DataLoader[DataId, DataInst], state: GEPAState, target_program_idx: ProgramIdx | None = None) -> list[DataId]

Always return the full ordered list of validation ids.

Source code in gepa/strategies/eval_policy.py 
37
38
39
40
41
def get_eval_batch(
    self, loader: DataLoader[DataId, DataInst], state: GEPAState, target_program_idx: ProgramIdx | None = None
) -> list[DataId]:
    """Always return the full ordered list of validation ids."""
    return list(loader.all_ids())

get_best_program(state: GEPAState) -> ProgramIdx

Pick the program whose evaluated validation scores achieve the highest average.

Source code in gepa/strategies/eval_policy.py 
43
44
45
46
47
48
49
50
51
52
53
def get_best_program(self, state: GEPAState) -> ProgramIdx:
    """Pick the program whose evaluated validation scores achieve the highest average."""
    best_idx, best_score, best_coverage = -1, float("-inf"), -1
    for program_idx, scores in enumerate(state.prog_candidate_val_subscores):
        coverage = len(scores)
        avg = sum(scores.values()) / coverage if coverage else float("-inf")
        if avg > best_score or (avg == best_score and coverage > best_coverage):
            best_score = avg
            best_idx = program_idx
            best_coverage = coverage
    return best_idx

get_valset_score(program_idx: ProgramIdx, state: GEPAState) -> float

Return the score of the program on the valset

Source code in gepa/strategies/eval_policy.py 
55
56
57
def get_valset_score(self, program_idx: ProgramIdx, state: GEPAState) -> float:
    """Return the score of the program on the valset"""
    return state.get_program_average_val_subset(program_idx)[0]