gpt-5.6-sol

{
  "input": "const r = await tools.exec_command({cmd:\"python -u - <<'PY'\\nimport numpy as np\\nfrom tokenizers import Tokenizer\\np='/root/.cache/huggingface/hub/models--gpt2/snapshots/607a30d783dfa663caf39e06633721c8d4cfcd7e/tokenizer.json'\\nt=Tokenizer.from_file(p); a=np.load('data/multi_dev.npy')\\ns=t.decode(a.tolist(),skip_special_tokens=False)\\nprint(len(s),s.count('<|endoftext|>'))\\nopen('/tmp/dev.txt','w').write(s)\\nPY\\nwc -c /tmp/dev.txt\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":1000});\ntext(r.output);"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Add File: /tmp/make_candidates.py\\n+import json, re, math, random, os\\n+from collections import Counter\\n+import numpy as np\\n+from tokenizers import Tokenizer\\n+from sklearn.feature_extraction.text import TfidfVectorizer\\n+from sklearn.linear_model import LogisticRegression\\n+\\n+POOL='/workspace/data/pool.jsonl'\\n+DEV='/workspace/data/multi_dev.npy'\\n+OUT='/workspace/candidates'\\n+os.makedirs(OUT,exist_ok=True)\\n+tp='/root/.cache/huggingface/hub/models--gpt2/snapshots/607a30d783dfa663caf39e06633721c8d4cfcd7e/tokenizer.json'\\n+tok=Tokenizer.from_file(tp)\\n+\\n+def chunks(s, n=1800):\\n+    # Keep examples reasonably homogeneous and discard tiny page fragments.\\n+    out=[]\\n+    for d in s.split('<|endoftext|>'):\\n+        d=d.strip()\\n+        if len(d)<300: continue\\n+        for i in range(0,len(d),n):\\n+            z=d[i:i+n]\\n+            if len(z)>=300: out.append(z)\\n+    return out\\n+\\n+a=np.load(DEV)\\n+dev_groups=[]\\n+for k in range(4):\\n+    s=tok.decode(a[k*250000:(k+1)*250000].tolist(),skip_special_tokens=False)\\n+    dev_groups.append(chunks(s))\\n+print('dev examples',list(map(len,dev_groups)),flush=True)\\n+\\n+# Deterministic broad sample of the pool for contrastive quality learning.\\n+neg=[]\\n+for i,line in enumerate(open(POOL)):\\n+    if i%29==0:\\n+        t=json.loads(line)['text'].strip()\\n+        if len(t)>=300: neg.append(t[:3600])\\n+print('negative examples',len(neg),flush=True)\\n+\\n+pos=[x for g in dev_groups for x in g]\\n+# Word/phrase features deliberately favor content and register, rather than exact URLs or IDs.\\n+vec=TfidfVectorizer(lowercase=True, strip_accents='unicode', analyzer='word',\\n+                    ngram_range=(1,2), min_df=2, max_df=.995, max_features=90000,\\n+                    sublinear_tf=True, token_pattern=r\\\"(?u)\\\\b[\\\\w][\\\\w+'-]+\\\\b\\\")\\n+X=vec.fit_transform(pos+neg)\\n+y=np.r_[np.ones(len(pos)),np.zeros(len(neg))]\\n+qclf=LogisticRegression(C=1.3,max_iter=150,class_weight='balanced',solver='liblinear').fit(X,y)\\n+\\n+# Domain model is trained on the four disclosed, equal-token target regions.\\n+dtexts=[]; dy=[]\\n+for k,g in enumerate(dev_groups):\\n+    dtexts.extend(g); dy.extend([k]*len(g))\\n+Xd=vec.transform(dtexts)\\n+dclf=LogisticRegression(C=3.0,max_iter=200,class_weight='balanced',solver='liblinear',multi_class='ovr').fit(Xd,dy)\\n+\\n+bad_re=re.compile(r'\\\\b(cookie|cookies|privacy policy|terms of use|sign in|log in|shopping cart|add to cart|skip to content|navigation|javascript|subscribe|newsletter|all rights reserved|password|username|contact us|site map|advertisement)\\\\b',re.I)\\n+spam_re=re.compile(r'\\\\b(viagra|cialis|casino|payday loan|essay writing service|porn|slots online|weight loss pills|buy cheap)\\\\b',re.I)\\n+sent_re=re.compile(r'[.!?][\\\\\\\"\\\\')\\\\]]?(?:\\\\s|$)')\\n+\\n+ids=[]; chars=[]; qs=[]; probs=[]; heur=[]\\n+batch=[]; bids=[]\\n+def score_batch(batch,bids):\\n+    Z=vec.transform(batch)\\n+    q=qclf.decision_function(Z)\\n+    p=dclf.predict_proba(Z)\\n+    for t,i,qq,pp in zip(batch,bids,q,p):\\n+        n=max(1,len(t)); words=re.findall(r\\\"[A-Za-z]+(?:'[A-Za-z]+)?\\\",t)\\n+        nw=max(1,len(words)); lines=[z.strip() for z in t.splitlines() if z.strip()]\\n+        bad=len(bad_re.findall(t)); spam=len(spam_re.findall(t))\\n+        alpha=sum(c.isalpha() for c in t)/n\\n+        # Signals of continuous prose and extraction cleanliness.\\n+        sentence=len(sent_re.findall(t))\\n+        prose=min(sentence/(nw/18+1),1.3)\\n+        short_lines=sum(len(z)<45 for z in lines)/max(1,len(lines))\\n+        repeated=1-len(set(lines))/max(1,len(lines))\\n+        eos=t.count('<|endoftext|>')\\n+        html=(t.count('<div')+t.count('<li')+t.count('<script')+t.count('{')*.15)/max(1,n/1000)\\n+        length_bonus=min(1.0,math.log1p(n)/8.2) - max(0,n-30000)/60000\\n+        h=(1.25*prose + 1.0*alpha + .45*length_bonus\\n+           - .11*bad/max(1,nw/250) - .8*spam\\n+           - .7*repeated - .38*short_lines - .08*max(0,eos-1) - .035*html)\\n+        ids.append(i); chars.append(n); qs.append(float(qq)); probs.append(pp); heur.append(h)\\n+\\n+for line in open(POOL):\\n+    o=json.loads(line); batch.append(o['text']); bids.append(o['id'])\\n+    if len(batch)==1000:\\n+        score_batch(batch,bids); batch=[]; bids=[]\\n+        if len(ids)%20000==0: print('scored',len(ids),flush=True)\\n+if batch: score_batch(batch,bids)\\n+ids=np.asarray(ids); chars=np.asarray(chars); qs=np.asarray(qs); probs=np.asarray(probs); heur=np.asarray(heur)\\n+np.savez('/workspace/candidates/scores.npz',ids=ids,chars=chars,q=qs,p=probs,h=heur)\\n+print('quality quantiles',np.quantile(qs,[0,.1,.5,.9,.99,1]),flush=True)\\n+\\n+def write(name,order):\\n+    order=list(map(int,order)); json.dump(order,open(f'{OUT}/{name}.json','w'))\\n+    print(name,len(order),'est M tok',round(chars[order].sum()/4.1/1e6,1),flush=True)\\n+\\n+rng=np.random.default_rng(1337)\\n+write('random',rng.permutation(len(ids))[:40000])\\n+\\n+# Basic filters rule out tiny snippets, giant scraped pages, and highly non-prose pages.\\n+valid=(chars>=500)&(chars<=35000)&(heur>0.15)\\n+source_prior=np.where(ids<115000,0.18,-0.10) # extraction regime shift visible in corpus statistics\\n+\\n+for tag,wq,wh,sp in [('q',1,0,0),('qh',1,.65,0),('qhsp',1,.65,1)]:\\n+    sc=wq*qs+wh*heur+sp*source_prior\\n+    ix=np.where(valid)[0]; ix=ix[np.argsort(-sc[ix],kind='stable')]\\n+    write(tag,ix[:50000])\\n+\\n+# Quota scheduler: rank separately for each target domain, then always draw from\\n+# the domain furthest below its 25% estimated-token share. A used set prevents duplicates.\\n+base=qs+.60*heur+source_prior\\n+rank=[]\\n+for d in range(4):\\n+    sc=base+1.15*np.log(np.maximum(probs[:,d],1e-5))\\n+    ix=np.where(valid)[0]; rank.append(ix[np.argsort(-sc[ix],kind='stable')])\\n+ptr=[0]*4; used=set(); totals=np.zeros(4); out=[]\\n+while totals.sum()<35_000_000*4.1 and len(out)<70000:\\n+    d=int(np.argmin(totals+1e-6*np.arange(4)))\\n+    while ptr[d]<len(rank[d]) and int(rank[d][ptr[d]]) in used: ptr[d]+=1\\n+    if ptr[d]>=len(rank[d]): break\\n+    j=int(rank[d][ptr[d]]); ptr[d]+=1; used.add(j); out.append(j); totals[d]+=chars[j]\\n+write('balanced',out)\\n+print('balanced chars/domain',totals,flush=True)\\n+\\n+# Softer quota candidate, useful if hard domain discrimination is noisy.\\n+for domweight in [.45,.75]:\\n+    rank=[]\\n+    for d in range(4):\\n+        sc=base+domweight*np.log(np.maximum(probs[:,d],1e-5))\\n+        ix=np.where(valid)[0]; rank.append(ix[np.argsort(-sc[ix],kind='stable')])\\n+    ptr=[0]*4; used=set(); totals=np.zeros(4); out=[]\\n+    while totals.sum()<35_000_000*4.1 and len(out)<70000:\\n+        d=int(np.argmin(totals+1e-6*np.arange(4)))\\n+        while ptr[d]<len(rank[d]) and int(rank[d][ptr[d]]) in used: ptr[d]+=1\\n+        if ptr[d]>=len(rank[d]): break\\n+        j=int(rank[d][ptr[d]]); ptr[d]+=1; used.add(j); out.append(j); totals[d]+=chars[j]\\n+    write('balanced'+str(domweight),out)\\n*** End Patch\";\nconst result = await tools.apply_patch(patch);\ntext(result);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"ls -l /tmp/make_candidates.py && env USE_TF=0 USE_FLAX=0 python -u /tmp/make_candidates.py\",\"workdir\":\"/workspace\",\"yield_time_ms\":1000,\"max_output_tokens\":3000,\"tty\":true}); text(JSON.stringify(r));\n"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Update File: /tmp/make_candidates.py\\n@@\\n-from sklearn.feature_extraction.text import TfidfVectorizer\\n-from sklearn.linear_model import LogisticRegression\\n@@\\n-pos=[x for g in dev_groups for x in g]\\n-# Word/phrase features deliberately favor content and register, rather than exact URLs or IDs.\\n-vec=TfidfVectorizer(lowercase=True, strip_accents='unicode', analyzer='word',\\n-                    ngram_range=(1,2), min_df=2, max_df=.995, max_features=90000,\\n-                    sublinear_tf=True, token_pattern=r\\\"(?u)\\\\b[\\\\w][\\\\w+'-]+\\\\b\\\")\\n-X=vec.fit_transform(pos+neg)\\n-y=np.r_[np.ones(len(pos)),np.zeros(len(neg))]\\n-qclf=LogisticRegression(C=1.3,max_iter=150,class_weight='balanced',solver='liblinear').fit(X,y)\\n-\\n-# Domain model is trained on the four disclosed, equal-token target regions.\\n-dtexts=[]; dy=[]\\n-for k,g in enumerate(dev_groups):\\n-    dtexts.extend(g); dy.extend([k]*len(g))\\n-Xd=vec.transform(dtexts)\\n-dclf=LogisticRegression(C=3.0,max_iter=200,class_weight='balanced',solver='liblinear',multi_class='ovr').fit(Xd,dy)\\n+word_re=re.compile(r\\\"[a-z][a-z+'-]{1,30}\\\")\\n+def feats(t):\\n+    w=word_re.findall(t.lower())\\n+    # Unigrams carry most of the stable register signal. Phrase features are\\n+    # prefixed so they cannot collide with words.\\n+    return w + ['~'+a+'_'+b for a,b in zip(w,w[1:])]\\n+\\n+def feature_counts(texts):\\n+    c=Counter(); total=0\\n+    for t in texts:\\n+        f=feats(t); c.update(f); total+=len(f)\\n+    return c,total\\n+\\n+pos=[x for g in dev_groups for x in g]\\n+pc,pt=feature_counts(pos); nc,nt=feature_counts(neg)\\n+vocab={x for x,n in pc.items() if n>=3}\\n+vocab.update(x for x,n in nc.items() if n>=5)\\n+V=len(vocab); alpha=.3\\n+qweight={x: max(-4,min(4,math.log((pc[x]+alpha)/(pt+alpha*V))-math.log((nc[x]+alpha)/(nt+alpha*V)))) for x in vocab}\\n+\\n+# Four equal-prior multinomial reference models for domain affinity.\\n+dcounts=[]; dtot=[]\\n+for g in dev_groups:\\n+    c,n=feature_counts(g); dcounts.append(c); dtot.append(n)\\n+dlog=[]\\n+for d in range(4):\\n+    den=dtot[d]+alpha*V\\n+    dlog.append({x:math.log((dcounts[d][x]+alpha)/den) for x in vocab})\\n+print('vocab',V,'features',pt,nt,flush=True)\\n@@\\n def score_batch(batch,bids):\\n-    Z=vec.transform(batch)\\n-    q=qclf.decision_function(Z)\\n-    p=dclf.predict_proba(Z)\\n-    for t,i,qq,pp in zip(batch,bids,q,p):\\n+    for t,i in zip(batch,bids):\\n+        fs=feats(t); fn=max(1,len(fs))\\n+        # Capping repeats makes menus and keyword stuffing less able to dominate.\\n+        fc=Counter(fs)\\n+        qq=sum(qweight.get(x,0.0)*min(n,4) for x,n in fc.items()) / max(1,sum(min(n,4) for n in fc.values()))\\n+        ds=np.array([sum(dlog[d].get(x,-15.0)*min(n,4) for x,n in fc.items())/max(1,sum(min(n,4) for n in fc.values())) for d in range(4)])\\n+        z=np.exp((ds-ds.max())*3.0); pp=z/z.sum()\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"env USE_TF=0 USE_FLAX=0 python -u /tmp/make_candidates.py\",\"workdir\":\"/workspace\",\"yield_time_ms\":1000,\"max_output_tokens\":3000,\"tty\":true}); text(JSON.stringify(r));\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:81373,chars:\"\",yield_time_ms:30000,max_output_tokens:4000}); text(r.output); text(JSON.stringify({session_id:r.session_id,exit_code:r.exit_code,wall:r.wall_time_seconds}));\n"
}

{
  "cell_id": "23",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:81373,chars:\"\",yield_time_ms:30000,max_output_tokens:4000}); text(r.output); text(JSON.stringify({session_id:r.session_id,exit_code:r.exit_code,wall:r.wall_time_seconds}));\n"
}

{
  "cell_id": "24",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:81373,chars:\"\",yield_time_ms:30000,max_output_tokens:4000}); text(r.output); text(JSON.stringify({session_id:r.session_id,exit_code:r.exit_code,wall:r.wall_time_seconds}));\n"
}

{
  "cell_id": "25",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:81373,chars:\"\",yield_time_ms:30000,max_output_tokens:4000}); text(r.output); text(JSON.stringify({session_id:r.session_id,exit_code:r.exit_code,wall:r.wall_time_seconds}));\n"
}

{
  "cell_id": "26",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:81373,chars:\"\",yield_time_ms:30000,max_output_tokens:4000}); text(r.output); text(JSON.stringify({session_id:r.session_id,exit_code:r.exit_code,wall:r.wall_time_seconds}));\n"
}

{
  "cell_id": "27",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:81373,chars:\"\",yield_time_ms:30000,max_output_tokens:4000}); text(r.output); text(JSON.stringify({session_id:r.session_id,exit_code:r.exit_code,wall:r.wall_time_seconds}));\n"
}

{
  "cell_id": "28",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:81373,chars:\"\",yield_time_ms:30000,max_output_tokens:4000}); text(r.output); text(JSON.stringify({session_id:r.session_id,exit_code:r.exit_code,wall:r.wall_time_seconds}));\n"
}

{
  "cell_id": "29",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:81373,chars:\"\",yield_time_ms:30000,max_output_tokens:4000}); text(r.output); text(JSON.stringify({session_id:r.session_id,exit_code:r.exit_code,wall:r.wall_time_seconds}));\n"
}

{
  "cell_id": "30",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:81373,chars:\"\",yield_time_ms:30000,max_output_tokens:4000}); text(r.output); text(JSON.stringify({session_id:r.session_id,exit_code:r.exit_code,wall:r.wall_time_seconds}));\n"
}

{
  "cell_id": "31",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}