Design a Complete Multimodal RLVR Pipeline with Open-MM-RL, Vision-Language Prompting, Reward Scoring, and GRPO Export

EXTRACT_PATS = [
   r"\\boxed\{([^{}]+)\}",
   r"final\s+answer\s*[:=]\s*([^\n]+)",
   r"answer\s*[:=]\s*([^\n]+)",
]
def extract_final(text):
   if not text: return ""
   for p in EXTRACT_PATS:
       m = re.search(p, text, flags=re.IGNORECASE)
       if m: return m.group(1).strip().strip(".,;")
   lines = [l.strip() for l in str(text).strip().splitlines() if l.strip()]
   return lines[-1] if lines else ""
def latex_to_sympy(s):
   s = (s or "").strip().strip("$").strip()
   s = re.sub(r"^\\[\[\(]", "", s); s = re.sub(r"\\[\]\)]$", "", s)
   s = (s.replace("\\pi", "pi").replace("\\cdot", "*").replace("\\times", "*")
          .replace("\\,", "").replace("\\;", "").replace("\\!", ""))
   s = re.sub(r"\\frac\s*\{([^{}]+)\}\s*\{([^{}]+)\}", r"((\1)/(\2))", s)
   s = re.sub(r"\\sqrt\s*\{([^{}]+)\}", r"sqrt(\1)", s)
   s = s.replace("^", "**")
   s = re.sub(r"\\[a-zA-Z]+", "", s)
   s = s.replace("{", "(").replace("}", ")")
   return s
def grade(pred, gold, tol=1e-4):
   """Verifiable reward in [0,1]: exact > numeric > sympy-symbolic > partial."""
   if pred is None or gold is None: return 0.0
   p = extract_final(str(pred)).strip()
   g = str(gold).strip()
   norm = lambda x: re.sub(r"\s+", "", x.lower()).strip("$.,;[]()")
   if norm(p) == norm(g): return 1.0
   def to_float(x):
       try: return float(latex_to_sympy(x))
       except Exception:
           try: return float(sp.sympify(latex_to_sympy(x)).evalf())
           except Exception: return None
   fp, fg = to_float(p), to_float(g)
   if fp is not None and fg is not None:
       if abs(fp - fg) / max(1.0, abs(fg)) < tol: return 1.0
   try:
       ep = sp.sympify(latex_to_sympy(p)); eg = sp.sympify(latex_to_sympy(g))
       if sp.simplify(ep - eg) == 0: return 1.0
   except Exception:
       pass
   if norm(g) and norm(g) in norm(p): return 0.5
   return 0.0
print("\n=== Grader sanity checks ===")
for pred, gold, want in [
   ("The answer is \\boxed{120}",            "[120]",            1.0),
   ("After computing: 7396 \\pi",            "7396\\pi",         1.0),
   ("Final answer: -71/4",                   "-\\frac{71}{4}",   1.0),
   ("Therefore the result is 0.0074",        "0.0074",           1.0),
   ("Final answer: nucleus accumbens",       "Nucleus accumbens",1.0),
   ("I don't know",                          "12",               0.0),
]:
   print(f"  pred={pred[:38]!r:42s} gold={gold!r:22s} -> r={grade(pred, gold)}  (want {want})")
SYSTEM = ("You are a STEM expert solving multimodal reasoning problems. "
         "You will see a question and one or more figures. "
         "Reason step by step, then end with exactly one line:\n"
         "Final answer: <your answer>")
def build_prompt(ex):
   img_tags = "\n".join(f"[Image {i+1}]" for i in range(len(ex["images"])))
   return f"{SYSTEM}\n\n{img_tags}\n\nQuestion:\n{ex['question']}\n\nLet's think step by step."
print("\n=== Example prompt (truncated) ===")
print(build_prompt(ds[0])[:600], "...\n")