scripts/temp_cal.py

import json, math, random, torch, sys
sys.path.insert(0,"/root/alea")
from alea.eval import load_model, predict_all
from transformers import AutoTokenizer
ckpt=sys.argv[1]; data=sys.argv[2]; out=sys.argv[3]; n=int(sys.argv[4]) if len(sys.argv)>4 else 4000
dev=torch.device("cuda")
tok=AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
model=load_model(ckpt,"Qwen/Qwen3-0.6B",dev)
rows=[json.loads(l) for l in open(data)]
random.seed(0); random.shuffle(rows); rows=rows[:n]
preds=predict_all(model,tok,rows,dev,max_len=4096)
# collect (probs, target) per type
ch=[]; sc=[]; nu=[]
for r in rows:
    pr=preds.get(r["id"],{})
    for qid,spec in r["questions"].items():
        t=r["targets"].get(qid); a=pr.get(qid)
        if t is None or a is None: continue
        if spec["type"]=="choice" and a.get("type")=="choice":
            ch.append((torch.tensor(a["probs"],dtype=torch.float64),torch.tensor(t,dtype=torch.float64)))
        elif spec["type"]=="score" and a.get("type")=="choice" :
            sc.append((torch.tensor(a["probs"],dtype=torch.float64),torch.tensor(t,dtype=torch.float64)))
        elif spec["type"]=="noul" and a.get("type")=="noul":
            nu.append((a["p"],float(t)))
def scale_probs(p,tau):
    q=torch.log(p.clamp_min(1e-12))/tau
    return torch.softmax(q,-1)
def brier_list(items,tau):
    s=0
    for p,y in items:
        q=scale_probs(p,tau); s+=((q-y)**2).sum().item()
    return s/len(items)
def noul_brier(items,tau):
    s=0
    for p,y in items:
        lp=math.log(max(p,1e-9)/(1-max(p,1e-9)))/tau
        q=1/(1+math.exp(-lp)); s+=(q-y)**2
    return s/len(items)
best={}
for name,items,fn in [("choice",ch,brier_list),("score",sc,brier_list),("noul",nu,noul_brier)]:
    if not items: best[name]=(1.0,0.0); continue
    cands=[x/20 for x in range(10,81)]
    scores=[(t,fn(items,t)) for t in cands]
    best[name]=min(scores,key=lambda z:z[1])
    base=fn(items,1.0)
    print(f"{name}: n={len(items)} base_brier={base:.4f} tau={best[name][0]:.2f} brier={best[name][1]:.4f}")
# bake
c=torch.load(ckpt,map_location="cpu")
tc=best["choice"][0]; ts=best["score"][0]; tn=best["noul"][0]
for k in ["out.weight","out.bias"]:
    c["heads"]["choice"][k]=c["heads"]["choice"][k]/tc
    c["heads"]["score"][k]=c["heads"]["score"][k]/ts
for k in ["net.3.weight","net.3.bias"]:
    c["heads"]["noul"][k]=c["heads"]["noul"][k]/tn
c["temp_cal"]={"choice":tc,"score":ts,"noul":tn,"fit_data":data,"n":n}
torch.save(c,out); print("saved",out)