Design a Full Multimodal RLVR Pipeline with Open-MM-RL, Imaginative and prescient-Language Prompting, Reward Scoring, and GRPO Export

EXTRACT_PATS = [
   r"boxed{([^{}]+)}",
   r"finals+solutions*[:=]s*([^n]+)",
   r"solutions*[:=]s*([^n]+)",
]
def extract_final(textual content):
   if not textual content: return ""
   for p in EXTRACT_PATS:
       m = re.search(p, textual content, flags=re.IGNORECASE)
       if m: return m.group(1).strip().strip(".,;")
   strains = [l.strip() for l in str(text).strip().splitlines() if l.strip()]
   return strains[-1] if strains else ""
def latex_to_sympy(s):
   s = (s or "").strip().strip("$").strip()
   s = re.sub(r"^[[(]", "", s); s = re.sub(r"[])]$", "", s)
   s = (s.exchange("pi", "pi").exchange("cdot", "*").exchange("instances", "*")
          .exchange(",", "").exchange(";", "").exchange("!", ""))
   s = re.sub(r"fracs*{([^{}]+)}s*{([^{}]+)}", r"((1)/(2))", s)
   s = re.sub(r"sqrts*{([^{}]+)}", r"sqrt(1)", s)
   s = s.exchange("^", "**")
   s = re.sub(r"[a-zA-Z]+", "", s)
   s = s.exchange("{", "(").exchange("}", ")")
   return s
def grade(pred, gold, tol=1e-4):
   """Verifiable reward in [0,1]: actual > 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.decrease()).strip("$.,;[]()")
   if norm(p) == norm(g): return 1.0
   def to_float(x):
       attempt: return float(latex_to_sympy(x))
       besides Exception:
           attempt: return float(sp.sympify(latex_to_sympy(x)).evalf())
           besides Exception: return None
   fp, fg = to_float(p), to_float(g)
   if fp will not be None and fg will not be None:
       if abs(fp - fg) / max(1.0, abs(fg)) < tol: return 1.0
   attempt:
       ep = sp.sympify(latex_to_sympy(p)); eg = sp.sympify(latex_to_sympy(g))
       if sp.simplify(ep - eg) == 0: return 1.0
   besides Exception:
       cross
   if norm(g) and norm(g) in norm(p): return 0.5
   return 0.0
print("n=== Grader sanity checks ===")
for pred, gold, need in [
   ("The answer is boxed{120}",            "[120]",            1.0),
   ("After computing: 7396 pi",            "7396pi",         1.0),
   ("Closing reply: -71/4",                   "-frac{71}{4}",   1.0),
   ("Subsequently the result's 0.0074",        "0.0074",           1.0),
   ("Closing reply: nucleus accumbens",       "Nucleus accumbens",1.0),
   ("I do not know",                          "12",               0.0),
]:
   print(f"  pred={pred[:38]!r:42s} gold={gold!r:22s} -> r={grade(pred, gold)}  (need {need})")
SYSTEM = ("You're a STEM knowledgeable fixing multimodal reasoning issues. "
         "You will note a query and a number of figures. "
         "Cause step-by-step, then finish with precisely one line:n"
         "Closing reply: ")
def build_prompt(ex):
   img_tags = "n".be part of(f"[Image {i+1}]" for i in vary(len(ex["images"])))
   return f"{SYSTEM}nn{img_tags}nnQuestion:n{ex['question']}nnLet's suppose step-by-step."
print("n=== Instance immediate (truncated) ===")
print(build_prompt(ds[0])[:600], "...n")