Reward Function Engineering: Verifiable, Math, Code, Format, Length, Diversity

# === Math reward — SymPy ile ===
import sympy
 
def math_reward_sympy(prompt, response, gold_expr):
    """SymPy ile expression equality — daha sağlam."""
    try:
        # Extract final answer
        import re
        match = re.search(r"####\s*([^\n]+)", response)
        if not match:
            return -1.0
 
        pred_str = match.group(1).strip()
        pred = sympy.sympify(pred_str)
        gold = sympy.sympify(gold_expr)
 
        # Numerical compare (also handles 2/4 == 0.5)
        if sympy.simplify(pred - gold) == 0:
            return 1.0
        return -0.5
    except:
        return -1.0   # parse error
 
# Examples:
# pred="0.5", gold="1/2" → reward = 1.0 (equal)
# pred="x+1", gold="1+x" → reward = 1.0 (algebraic equal)

# === Code execution reward ===
import subprocess, tempfile, os
 
def code_exec_reward(prompt, response, test_cases, timeout=5):
    """Response içindeki Python code'u extract et, test cases'le çalıştır."""
    # Extract code
    import re
    code_match = re.search(r"\\`\\`\\`python\n(.*?)\\`\\`\\`", response, re.DOTALL)
    if not code_match:
        return -1.0
    code = code_match.group(1)
 
    # Write + run + collect passes
    with tempfile.NamedTemporaryFile(suffix=".py", delete=False, mode="w") as f:
        f.write(code + "\n\n")
        for test in test_cases:
            f.write(test + "\n")
        tmp_path = f.name
 
    try:
        result = subprocess.run(
            ["python", tmp_path],
            capture_output=True, text=True, timeout=timeout,
        )
        passed = result.returncode == 0
    except subprocess.TimeoutExpired:
        passed = False
    finally:
        os.unlink(tmp_path)
 
    return 1.0 if passed else -0.5
 
# Sandbox: production'da Docker container içinde çalıştırılır
# Cookbook'un Part XVII'sinde detay