benak / code /thinking.py
PugarHuda's picture
Benak release
7debf49 verified
Raw History Blame Contribute Delete
3.54 kB
"""Babak 3: build in-language reasoning traces where Beam got the answer right but thought in English.
Beam rewrites its own correct reasoning into the target language (as visible content, which follows language
instructions far better than reasoning does). Kept only if GlotLID says >= 0.8 target language AND every number
of the original reasoning survives. Output: data/thinking/rewritten.jsonl (trace_source='rewritten').
Usage: python thinking.py --per-lang 60"""
import argparse, json, os, re
from beam import chat, run_many, spent
from lid import profile
LANGS = ["ind_Latn", "jav_Latn", "sun_Latn", "zsm_Latn", "tgl_Latn", "ceb_Latn", "vie_Latn", "tha_Thai"]
# v2: v1 let Beam *solve* the embedded problem (251/439 outputs were just \boxed{answer}); now the text is fenced
# and the task is stated as translation-only.
PROMPT = """You are a translator. Translate the text inside <notes> into natural, fluent {name}.
This is a translation task only: do NOT solve anything, do NOT add a final answer, do NOT shorten.
Keep every step, number and calculation exactly as written. Output only the translated text, without the tags.
<notes>
{reasoning}
</notes>"""
def nums(t):
t = re.sub(r"(?<=\d)[.,](?=\d{3}\b)", "", t) # drop thousands separators (1.250.000 / 1,250,000)
return set(re.findall(r"\d+", t)) # ponytail: digit-run set; tolerant to 1,250 vs 1250 formatting
def main():
ap = argparse.ArgumentParser(); ap.add_argument("--per-lang", type=int, default=60); a = ap.parse_args()
names = {c: v["lang_name_native"] for c, v in json.load(open("data/atlas/templates.json", encoding="utf-8")).items()}
par = json.load(open("data/math/parallel.json", encoding="utf-8"))
runs = [json.loads(l) for l in open("data/math/runs_low.jsonl", encoding="utf-8")]
pool = {}
for r in runs: # correct, reasoning mostly NOT in the target language
ts = r["lid"]["target_share"]
if r["lang"] in LANGS and r["correct"] and r["reasoning"] and (ts is None or ts < 0.8):
pool.setdefault(r["lang"], {}).setdefault(r["id"], r) # one trace per problem per language
jobs = [r for l in LANGS for r in list(pool.get(l, {}).values())[:a.per_lang]]
def one(r):
out = chat([{"role": "user", "content": PROMPT.format(name=names[r["lang"]], reasoning=r["reasoning"])}],
"low", 6000, temperature=0.2, tag="rewrite-v2")
new = out["content"].strip()
lid = profile(new, r["lang"])
a, b = nums(new), nums(r["reasoning"])
jac = len(a & b) / max(1, len(a | b)) # number-set Jaccard; translation may re-spell a few ordinals
ok = (lid["target_share"] or 0) >= 0.8 and jac >= 0.9
return {"lang": r["lang"], "id": r["id"], "answer": r["answer"], "effort": r["effort"],
"user": par[str(r["id"])][r["lang"]]["text"], "reasoning": new, "content": r["content"],
"correct": ok, "lid": lid, "numbers_jaccard": round(jac, 3),
"original_reasoning_lid": r["lid"], "trace_source": "rewritten"}
rows = [x for _, x in run_many(one, jobs, 10, "rewrite") if not isinstance(x, Exception)]
os.makedirs("data/thinking", exist_ok=True)
with open("data/thinking/rewritten.jsonl", "w", encoding="utf-8") as f:
f.writelines(json.dumps(x, ensure_ascii=False) + "\n" for x in rows)
print(f"{sum(x['correct'] for x in rows)}/{len(rows)} rewrites kept; spent ${spent()}")
if __name__ == "__main__":
main()