pip install gepa

import gepa

# Load AIME datasets
trainset, valset, testset = gepa.examples.aime.init_dataset()

print(f"Training examples: {len(trainset)}")
print(f"Validation examples: {len(valset)}")
print(f"Test examples: {len(testset)}")

seed_prompt = {
    "system_prompt": "You are a helpful assistant. Answer the question. "
                     "Put your final answer in the format '### <answer>'"
}

result = gepa.optimize(
    seed_candidate=seed_prompt,
    trainset=trainset,
    valset=valset,
    task_lm="openai/gpt-4.1-mini",
    max_metric_calls=150,
    reflection_lm="openai/gpt-5",
)

print("Optimized prompt:", result.best_candidate['system_prompt'])
print(f"Best validation score: {result.val_aggregate_scores[result.best_idx]:.2%}")

You will be given one math problem as plain text. Your job is to solve it 
correctly and return:

- reasoning: a concise, logically ordered solution that uses identities/structure 
  to avoid brute force, ends with a quick verification.
- answer: the final requested number/expression only (no extra words).

Domain-specific strategies:

1) Base-conversion/digit rearrangement:
   - Translate positional notation correctly
   - Enforce digit ranges strictly
   - Use modular constraints to prune (e.g., mod 9 often collapses coefficients)
   - Solve within digit bounds and verify numerically

2) Palindromes across bases:
   - Bound the base length by magnitude
   - Characterize palindromes algebraically
   - For "greatest", check candidates in descending order

3) Symmetric sums with fixed constraints:
   - Use identities to compress expressions
   - Convert relations among ab+bc+ca and abc
   - Count ordered solutions carefully

4) Intersecting families of subsets:
   - Empty set cannot be included
   - Complement pairs cannot both be present
   - Use size-based pigeonhole arguments

Quality checks:
- Verify digit/base constraints numerically
- For extremal problems, provide both bound and construction
- For counting, handle ordered vs unordered cases
- Justify optimality structurally

from gepa.adapters.default_adapter import DefaultAdapter

# Create adapter for evaluation
adapter = DefaultAdapter(
    task_lm="openai/gpt-4.1-mini",
    metric=lambda pred, ref: int(pred.strip() == ref.strip())
)

# Test on AIME 2025
test_result = adapter.evaluate(
    batch=testset,
    candidate=result.best_candidate,
    capture_traces=False
)

test_accuracy = sum(test_result.scores) / len(test_result.scores)
print(f"AIME 2025 Test Accuracy: {test_accuracy:.2%}")