"""benak_reward: a verifiable GRPO reward for "think in the user's language" (TRL GRPOTrainer-compatible). reward = correctness * (0.5 + 0.5 * in_language_share_of_thinking) - correctness: 1 if the final \\boxed{} answer matches `answer` (numeric, tolerant to 1.250.000 / 1,250,000 / 12,5) - in_language_share: GlotLID share of the thinking text in `lang`, after dropping sentences that quote the prompt So a right answer reasoned in English earns 0.5, a right answer reasoned in the target language earns 1.0, and a wrong answer earns 0. This fixes Qi et al.'s accuracy-vs-language trade-off in the reward rather than the prompt. Thinking is read from ... in the completion, or from a separate `thinking` column if TRL passes one. trainer = GRPOTrainer(model=..., reward_funcs=[benak_reward], train_dataset=ds) # ds has `answer`, `lang`, `prompt` """ import re from lid import profile _THINK = re.compile(r"(.*?)", re.S) def _text(c): return c if isinstance(c, str) else c[-1]["content"] def _boxed(t): i = t.rfind("\\boxed{") if i < 0: return None i += 7; d = 1; j = i while j < len(t) and d: d += {"{": 1, "}": -1}.get(t[j], 0); j += 1 return t[i:j - 1] if d == 0 else None def _num(s): s = re.sub(r"(?i)rp\.?|\s|%|\\[a-z]+\{|\}", "", s or "") m = re.findall(r"-?\d[\d.,]*", s) if not m: return None t = m[-1].rstrip(".,") if "," in t and "." in t: t = t.replace(".", "").replace(",", ".") if t.rfind(",") > t.rfind(".") else t.replace(",", "") elif t.count(".") > 1 or t.count(",") > 1 or re.fullmatch(r"-?\d{1,3}[.,]\d{3}", t): t = t.replace(".", "").replace(",", "") else: t = t.replace(",", ".") try: return float(t) except ValueError: return None def score(completion, answer, lang, prompt="", thinking=None): text = _text(completion) th = thinking if thinking is not None else "\n".join(_THINK.findall(text)) final = _THINK.sub("", text) got = _num(_boxed(final) or "") if got is None or abs(got - float(answer)) > max(0.011, 1e-4 * abs(float(answer))): return 0.0 share = profile(th, lang, exclude=prompt or None)["target_share"] return round(0.5 + 0.5 * (share or 0.0), 4) def benak_reward(completions, answer, lang, prompts=None, thinking=None, **_): """TRL reward_funcs signature: one float per completion.""" n = len(completions) prompts = prompts or [""] * n thinking = thinking or [None] * n return [score(c, a, l, _text(p) if p else "", t) for c, a, l, p, t in zip(completions, answer, lang, prompts, thinking)] if __name__ == "__main__": idn = "Pak Budi membeli tiga karung beras, masing-masing dua puluh lima kilogram. Jadi totalnya tujuh puluh lima kilogram dikali harga per kilogram. Jadi \\boxed{937.500}" eng = "Mr Budi buys three sacks of rice, each twenty five kilograms. So the total is seventy five kilograms times the price per kilogram. So \\boxed{937500}" bad = "Pak Budi membeli tiga karung beras, masing-masing dua puluh lima kilogram dan seterusnya. \\boxed{900000}" r = benak_reward([idn, eng, bad], [937500, 937500, 937500], ["ind_Latn"] * 3) assert r[0] > 0.95 and abs(r[1] - 0.5) < 0.05 and r[2] == 0.0, r assert benak_reward([[{"role": "assistant", "content": idn}]], [937500], ["ind_Latn"])[0] > 0.95 assert _num("Rp1.250.000,50") == 1250000.5 and _num("12,5") == 12.5 and _num("1,250") == 1250.0 print("benak_reward self-check ok", r)