Blame for code/score.py

import json,re,collections,statistics as st
d=json.load(open("raw_results.json"))
TA_DIG=str.maketrans("௦௧௨௩௪௫௬௭௮௯","0123456789")
def norm(x): return re.sub(r"[\s\.\,\!\?\:\"'`*()\[\]₹]|rs|rupees|ரூபாய்","",(x or "").lower().translate(TA_DIG))
def correct(r):
    o=norm(r["output"]); golds=[norm(r["answer"])]+[norm(a) for a in r["aliases"]]
    if r["answer"].isdigit():
        nums=re.findall(r"\d+(?:\.\d+)?", (r["output"] or "").translate(TA_DIG).replace(",",""))
        return bool(nums) and float(nums[-1])==float(r["answer"])
    return any(g and (o==g or o.startswith(g) or g in o) for g in golds)
PIN,POUT=0.40,1.20
for r in d: r["correct"]=correct(r); r["cost_usd"]=(r["prompt_tokens"]*PIN+r["completion_tokens"]*POUT)/1e6
json.dump(d,open("scored_results.json","w"),ensure_ascii=False,indent=1)
def summ(rows):
    return dict(n=len(rows),acc=round(100*sum(r["correct"] for r in rows)/len(rows),1),
     prompt=round(st.mean(r["prompt_tokens"] for r in rows),1),
     reasoning=round(st.mean(r["reasoning_tokens"] or 0 for r in rows),1),
     completion=round(st.mean(r["completion_tokens"] for r in rows),1),
     cost_per_1k=round(1000*st.mean(r["cost_usd"] for r in rows),4),
     latency=round(st.mean(r["latency"] for r in rows),2))
S={}
for l in("en","ta","tl"):
    S[l]=summ([r for r in d if r["lang"]==l])
    for c in("math","logic","knowledge","everyday"):
        S[f"{l}/{c}"]=summ([r for r in d if r["lang"]==l and r["cat"]==c])
print(json.dumps(S,indent=0))
wrong=[(r["id"],r["lang"],r["answer"],(r["output"] or "")[:40]) for r in d if not r["correct"]]
print(len(wrong)); [print(w) for w in sorted(wrong)]
print("total cost $",round(sum(r["cost_usd"] for r in d),4))
json.dump(S,open("summary.json","w"),indent=1)