gpt-5.6-sol

{
  "input": "const r = await tools.exec_command({cmd:\"ps aux | rg 'pack_selection|train_nano|dev_eval' | rg -v rg || true; ls -lh /tmp/dev_train.npy /tmp/dev.json 2>/dev/null || true; tail -5 /proc/732/fd/1 2>/dev/null || true\",workdir:\"/workspace\",yield_time_ms:10000,max_output_tokens:3000});\ntext(r.output);"
}

{
  "input": "const r = await tools.exec_command({cmd:\"sleep 20; ps aux | rg 'pack_selection|train_nano|dev_eval' | rg -v rg || true; ls -lh /tmp/dev_train.npy /tmp/dev.json 2>/dev/null || true\",workdir:\"/workspace\",yield_time_ms:30000,max_output_tokens:3000});\ntext(r.output);"
}

{
  "cell_id": "20",
  "yield_time_ms": 20000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.exec_command({cmd:\"sleep 25; ps -p 985 -o pid,etime,pcpu || true; cat /tmp/dev.json 2>/dev/null || true; cat /workspace/.timer/remaining_secs\",workdir:\"/workspace\",yield_time_ms:30000,max_output_tokens:3000});\ntext(r.output);"
}

{
  "cell_id": "21",
  "yield_time_ms": 20000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.exec_command({cmd:\"sleep 25; ps -p 985 -o pid,etime,pcpu || true; cat /tmp/dev.json 2>/dev/null || true\",workdir:\"/workspace\",yield_time_ms:30000,max_output_tokens:3000});\ntext(JSON.stringify(r));"
}

{
  "cell_id": "22",
  "yield_time_ms": 20000,
  "max_tokens": 3000
}

{
  "input": "let{output,...rest}=await tools.exec_command({cmd:\"cp /tmp/dev.json /workspace/result_random.json && bash /workspace/dev_eval.sh /workspace/first110.json\",workdir:\"/workspace\",yield_time_ms:30000,max_output_tokens:3000});text(rest);text(output);"
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps aux | rg 'pack_selection|train_nano' | rg -v rg || true; cat /tmp/dev.json 2>/dev/null; cat /workspace/.timer/remaining_secs\",workdir:\"/workspace\",yield_time_ms:10000,max_output_tokens:3000});\ntext(r.output);"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Add File: /workspace/target_rank.py\\n+#!/usr/bin/env python3\\n+\\\"\\\"\\\"Prototype target-style classifier and balanced ranking.\\\"\\\"\\\"\\n+import html, json, random, re\\n+import numpy as np\\n+from transformers import GPT2TokenizerFast\\n+from sklearn.feature_extraction.text import TfidfVectorizer\\n+from sklearn.linear_model import SGDClassifier\\n+\\n+POOL = '/workspace/data/pool.jsonl'\\n+DEV = '/workspace/data/multi_dev.npy'\\n+OUT = '/workspace/target.json'\\n+TAG = re.compile(r'<[^>]{1,200}>')\\n+SPACE = re.compile(r'\\\\s+')\\n+BAD = ('cookie policy','privacy policy','terms of use','skip to content','sign in',\\n+       'log in','all rights reserved','javascript','shopping cart','free shipping',\\n+       'contact us','click here','subscribe','toggle navigation','search menu')\\n+\\n+def norm(x):\\n+    x = html.unescape(x.replace(' @-@ ', '-').replace(' @,@ ', ','))\\n+    x = TAG.sub(' ', x)\\n+    return SPACE.sub(' ', x).strip().lower()\\n+\\n+def quality(x, idx):\\n+    n=max(1,len(x)); lo=x.lower(); words=re.findall(r\\\"[A-Za-z]+(?:'[A-Za-z]+)?\\\",x)\\n+    nw=max(1,len(words)); lines=x.splitlines()\\n+    alpha=len(re.findall(r'[A-Za-z\\\\s]',x))/n\\n+    sent=sum(x.count(c) for c in '.?!')/nw\\n+    short=sum(len(z.strip())<40 for z in lines)/max(1,len(lines))\\n+    uniq=len(set(w.lower() for w in words))/nw\\n+    q=1.8*alpha + .8*min(sent/.06,1.0)+.5*min(uniq/.55,1.0)\\n+    q-=.28*min(8,sum(lo.count(z) for z in BAD))\\n+    q-=.8*max(0,short-.35)-.3*(n>1500)-.8*(n<500)\\n+    q-=.35*(idx>=110000)\\n+    return q\\n+\\n+tok=GPT2TokenizerFast.from_pretrained('gpt2',local_files_only=True)\\n+dev=np.load(DEV)\\n+train_x=[]; train_y=[]\\n+for d in range(4):\\n+    txt=tok.decode(dev[d*250000:(d+1)*250000])\\n+    docs=txt.split('<|endoftext|>')\\n+    for doc in docs:\\n+        z=norm(doc)\\n+        if len(z)>=180:\\n+            # Comparable sized examples prevent source document length becoming a label.\\n+            for st in range(0,min(len(z),8000),2500):\\n+                if len(z[st:st+3000])>=180:\\n+                    train_x.append(z[st:st+3000]); train_y.append(d)\\n+\\n+rng=random.Random(9173); neg=[]; rows=[]\\n+with open(POOL) as f:\\n+    for line in f:\\n+        r=json.loads(line); rows.append((r['id'],r['text']))\\n+        # deterministic reservoir of web negatives\\n+        if len(neg)<16000: neg.append(r['text'][:3000])\\n+        else:\\n+            j=rng.randrange(r['id']+1)\\n+            if j<len(neg): neg[j]=r['text'][:3000]\\n+train_x.extend(norm(x) for x in neg); train_y.extend([4]*len(neg))\\n+print('examples', {y:train_y.count(y) for y in range(5)},flush=True)\\n+vec=TfidfVectorizer(ngram_range=(1,2),min_df=2,max_df=.995,max_features=180000,\\n+                    sublinear_tf=True,strip_accents='unicode',dtype=np.float32)\\n+X=vec.fit_transform(train_x)\\n+clf=SGDClassifier(loss='log_loss',alpha=1.5e-6,max_iter=40,tol=1e-4,\\n+                  class_weight='balanced',random_state=719)\\n+clf.fit(X,train_y)\\n+del X,train_x,neg\\n+\\n+scored=[]\\n+for off in range(0,len(rows),2048):\\n+    rr=rows[off:off+2048]\\n+    X=vec.transform([norm(x[:12000]) for _,x in rr])\\n+    dec=clf.decision_function(X)\\n+    for (idx,x),s in zip(rr,dec):\\n+        d=int(np.argmax(s[:4])); rel=float(s[d]-s[4]); q=quality(x,idx)\\n+        # Relevance is capped: quality should be able to veto keyword-rich junk.\\n+        score=0.65*max(-5,min(5,rel))+q\\n+        scored.append((idx,d,score,rel,q,len(x)))\\n+    if off%20480==0: print('scored',off,flush=True)\\n+\\n+by=[[] for _ in range(4)]\\n+for z in scored: by[z[1]].append(z)\\n+for d in range(4): by[d].sort(key=lambda z:(-z[2],z[0]))\\n+# Fair interleaving by estimated token mass; actual packing remains safely over budget.\\n+pos=[0]*4; mass=[0.0]*4; result=[]\\n+while len(result)<50000 and any(pos[d]<len(by[d]) for d in range(4)):\\n+    d=min((x for x in range(4) if pos[x]<len(by[x])),key=lambda x:mass[x])\\n+    z=by[d][pos[d]]; pos[d]+=1; result.append(z[0]); mass[d]+=max(1,z[5]/4)\\n+with open(OUT,'w') as f: json.dump(result,f)\\n+print('mass',mass,'counts',pos)\\n+lookup=dict(rows)\\n+for d in range(4):\\n+    print('\\\\nDOMAIN',d)\\n+    for z in by[d][:8]: print(z,repr(lookup[z[0]][:220].replace('\\\\n',' ')))\\n*** End Patch\";\ntext(await tools.apply_patch(patch));"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python /workspace/target_rank.py\",workdir:\"/workspace\",yield_time_ms:1000,max_output_tokens:5000});\ntext(JSON.stringify(r));"
}

{
  "input": "const r = await tools.write_stdin({session_id:47471,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r.output);"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python - <<'PY'\\ntry: import scipy; print('scipy',scipy.__version__)\\nexcept Exception as e: print(e)\\ntry: import pandas; print('pandas')\\nexcept Exception as e: print(e)\\nPY\\nps aux | rg 'train_nano' | rg -v rg || true; cat /tmp/dev.json\",workdir:\"/workspace\",yield_time_ms:10000,max_output_tokens:3000});\ntext(r.output);"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Update File: /workspace/target_rank.py\\n@@\\n-import html, json, random, re\\n+import html, json, math, random, re\\n+from collections import Counter\\n import numpy as np\\n from transformers import GPT2TokenizerFast\\n-from sklearn.feature_extraction.text import TfidfVectorizer\\n-from sklearn.linear_model import SGDClassifier\\n@@\\n def norm(x):\\n@@\\n     return SPACE.sub(' ', x).strip().lower()\\n+\\n+def terms(x):\\n+    w=re.findall(r\\\"[a-z]+(?:'[a-z]+)?|[0-9]+\\\",norm(x))\\n+    return w, [a+' '+b for a,b in zip(w,w[1:])]\\n@@\\n-train_x=[]; train_y=[]\\n+target=[Counter() for _ in range(4)]\\n+target_n=[[0,0] for _ in range(4)]\\n for d in range(4):\\n@@\\n-                    train_x.append(z[st:st+3000]); train_y.append(d)\\n+                    w,b=terms(z[st:st+3000]); target[d].update(w); target[d].update(b)\\n+                    target_n[d][0]+=len(w); target_n[d][1]+=len(b)\\n@@\\n-train_x.extend(norm(x) for x in neg); train_y.extend([4]*len(neg))\\n-print('examples', {y:train_y.count(y) for y in range(5)},flush=True)\\n-vec=TfidfVectorizer(ngram_range=(1,2),min_df=2,max_df=.995,max_features=180000,\\n-                    sublinear_tf=True,strip_accents='unicode',dtype=np.float32)\\n-X=vec.fit_transform(train_x)\\n-clf=SGDClassifier(loss='log_loss',alpha=1.5e-6,max_iter=40,tol=1e-4,\\n-                  class_weight='balanced',random_state=719)\\n-clf.fit(X,train_y)\\n-del X,train_x,neg\\n+vocab=set()\\n+for c in target:\\n+    vocab.update(k for k,v in c.items() if v>=3)\\n+bg=Counter(); bg_n=[0,0]\\n+for x in neg:\\n+    w,b=terms(x); bg.update(z for z in w if z in vocab); bg.update(z for z in b if z in vocab)\\n+    bg_n[0]+=len(w); bg_n[1]+=len(b)\\n+weights=[]\\n+for d in range(4):\\n+    wd={}\\n+    # Per-token smoothed log frequency ratio. Clip rare proper-name effects.\\n+    for z in vocab:\\n+        bi=' ' in z; tn=target_n[d][1 if bi else 0]; bn=bg_n[1 if bi else 0]\\n+        a=target[d].get(z,0); b=bg.get(z,0)\\n+        val=math.log((a+.3)/(tn+1)) - math.log((b+.3)/(bn+1))\\n+        wd[z]=max(-3.0,min(3.0,val))\\n+    weights.append(wd)\\n+print('target sizes',target_n,'background',bg_n,'vocab',len(vocab),flush=True)\\n+del target,bg,neg,vocab\\n@@\\n-    X=vec.transform([norm(x[:12000]) for _,x in rr])\\n-    dec=clf.decision_function(X)\\n-    for (idx,x),s in zip(rr,dec):\\n-        d=int(np.argmax(s[:4])); rel=float(s[d]-s[4]); q=quality(x,idx)\\n+    for idx,x in rr:\\n+        w,b=terms(x[:12000]); den=max(30,len(w))\\n+        s=[]\\n+        for d in range(4):\\n+            # Bigrams carry source/register style without dominating unigrams.\\n+            s.append((sum(weights[d].get(z,0) for z in w)+\\n+                      .35*sum(weights[d].get(z,0) for z in b))/den)\\n+        d=int(np.argmax(s)); rel=float(s[d]); q=quality(x,idx)\\n@@\\n-        score=0.65*max(-5,min(5,rel))+q\\n+        score=1.8*max(-2,min(2,rel))+q\\n*** End Patch\";\ntext(await tools.apply_patch(patch));"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python /workspace/target_rank.py\",workdir:\"/workspace\",yield_time_ms:1000,max_output_tokens:5000});\ntext(JSON.stringify(r));"
}

{
  "input": "const r = await tools.write_stdin({session_id:38983,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r.output);"
}

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

{
  "input": "const r = await tools.write_stdin({session_id:38983,chars:\"\",yield_time_ms:30000,max_output_tokens:10000});\ntext(r.output);"
}

{
  "cell_id": "32",
  "yield_time_ms": 30000,
  "max_tokens": 10000
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps aux | rg 'train_nano|target_rank' | rg -v rg; cat /tmp/dev.json; cat /workspace/.timer/remaining_secs\",workdir:\"/workspace\",yield_time_ms:10000,max_output_tokens:3000});\ntext(r.output);"
}

{
  "input": "const r = await tools.write_stdin({session_id:38983,chars:\"\",yield_time_ms:30000,max_output_tokens:10000});\ntext(r.output);"
}

{
  "cell_id": "34",
  "yield_time_ms": 30000,
  "max_tokens": 10000
}

{
  "input": "const r = await tools.write_stdin({session_id:38983,chars:\"\",yield_time_ms:30000,max_output_tokens:10000});\ntext(r.output);"
}

{
  "cell_id": "35",
  "yield_time_ms": 30000,
  "max_tokens": 10000
}