gpt-5.6-sol

{
  "input": "const r = await tools.exec_command({cmd:\"ps -eo pid,etime,pcpu,rss,args | grep -E 'extra_scan|token_lengths' | grep -v grep\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":3000});\ntext(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python - <<'PY'\\nimport json,pickle,numpy as np\\nz=pickle.load(open('scores.pkl','rb')); p=z['probs']; ids=z['ids']; mp={int(x):p[i].argmax() for i,x in enumerate(ids)}\\nsel=json.load(open('experiments/bal010.json'))\\n# approximate first 10503 used from prior output\\nu=sel[:10503]\\nfor k in range(4):\\n a=[x for x in u if mp[x]==k]; print(k,len(a),sum(x>=110000 for x in a),np.quantile(a,[0,.25,.5,.75,1]).astype(int))\\nPY\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":3000});\ntext(r.output);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:90697,chars:\"\",yield_time_ms:30000,max_output_tokens:3000});\ntext(r.output);\n"
}

{
  "cell_id": "56",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.write_stdin({session_id:63932,chars:\"\",yield_time_ms:30000,max_output_tokens:3000});\ntext(r.output);\n"
}

{
  "cell_id": "57",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.write_stdin({session_id:90697,chars:\"\",yield_time_ms:30000,max_output_tokens:3000});\ntext(r.output);\n"
}

{
  "cell_id": "58",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Update File: /workspace/make_exp.py\\n@@\\n else:\\n     z=pickle.load(open('/workspace/scores.pkl','rb'))\\n     ids=z['ids']; q=z['score'].copy(); s=z['stats']; p=z['probs']\\n+    lens=np.load('/workspace/token_lengths.npy')\\n+    ex=pickle.load(open('/workspace/extra.pkl','rb'))\\n+    # id,len,tech,html,news,wiki,bad,codechars,weird,moj,exact_hash\\n+    exnum=np.asarray(ex[:,:10],dtype=float); hashes=ex[:,10]\\n@@\\n-    qual=q+prior\\n+    qual=q+prior\\n+    # Extra cleanliness terms are deliberately small relative to target-density.\\n+    charlen,tech,html,news,wiki,bad,codechars,weird,moj=exnum[:,1:].T\\n+    qual2=qual - .10*np.log1p(bad) - .035*moj - .035*np.maximum(codechars-5,0)\\n+    qual2 -= .10*np.maximum(np.log(np.maximum(lens,1)/8000),0)\\n@@\\n-    else:\\n+    else:\\n+        if mode.startswith('v2'):\\n+            qual=qual2\\n         # Candidate domain assignment by closest target quarter, then fair token-proxy scheduling.\\n         dom=p.argmax(1)\\n         lists=[]\\n         for k in range(4):\\n             dk=np.where(dom==k)[0]\\n             # Quality is primary; modest confidence bonus avoids obvious domain mistakes.\\n-            val=qual[dk]+float(mode.split('_')[-1] if '_' in mode else 0.15)*np.log(p[dk,k]+.03)\\n+            if mode.startswith('v2'):\\n+                bonus=[.06*np.log1p(wiki[dk]),np.zeros(len(dk)),.04*np.log1p(news[dk]),\\n+                       .05*np.log1p(tech[dk])+.025*np.log1p(html[dk])][k]\\n+                val=qual[dk]+.10*np.log(p[dk,k]+.03)+bonus\\n+            else:\\n+                val=qual[dk]+float(mode.split('_')[-1] if '_' in mode else 0.15)*np.log(p[dk,k]+.03)\\n             lists.append(dk[np.argsort(-val)].tolist())\\n-        ptr=[0]*4; used=set(); order=[]; load=[0.0]*4\\n-        # weighted fair queue by word count proxy\\n+        ptr=[0]*4; used=set(); used_hash=set(); order=[]; load=[0.0]*4\\n+        weights=[1,1,1,1]\\n+        if mode=='v2_tech085': weights=[1,1,1,.85]\\n+        if mode=='v2_tech115': weights=[1,1,1,1.15]\\n+        # Weighted fair queue using exact GPT-2 token lengths.\\n         while len(order)<len(ids):\\n-            k=min(range(4),key=lambda j:load[j])\\n-            while ptr[k]<len(lists[k]) and int(ids[lists[k][ptr[k]]]) in used: ptr[k]+=1\\n+            k=min(range(4),key=lambda j:load[j]/weights[j])\\n+            while ptr[k]<len(lists[k]) and (int(ids[lists[k][ptr[k]]]) in used or hashes[lists[k][ptr[k]]] in used_hash): ptr[k]+=1\\n             if ptr[k]>=len(lists[k]):\\n                 load[k]=1e30\\n                 if min(load)>1e20: break\\n                 continue\\n-            ix=lists[k][ptr[k]]; ptr[k]+=1; did=int(ids[ix]); used.add(did); order.append(ix)\\n-            load[k]+=max(s[ix,7],1)\\n+            ix=lists[k][ptr[k]]; ptr[k]+=1; did=int(ids[ix]); used.add(did); used_hash.add(hashes[ix]); order.append(ix)\\n+            load[k]+=lens[did]\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python make_exp.py balexact experiments/balexact.json && python make_exp.py v2_equal experiments/v2eq.json && python make_exp.py v2_tech085 experiments/v2t085.json && python make_exp.py v2_tech115 experiments/v2t115.json && bash dev_eval.sh experiments/v2eq.json\",\"workdir\":\"/workspace\",\"yield_time_ms\":1000,\"max_output_tokens\":10000});\ntext(r.output);\nif (r.session_id) text(`SESSION_ID=${r.session_id}`);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:25923,chars:\"\",yield_time_ms:30000,max_output_tokens:10000});\ntext(r.output);\n"
}

{
  "cell_id": "61",
  "yield_time_ms": 30000,
  "max_tokens": 10000
}

{
  "input": "const r = await tools.write_stdin({session_id:25923,chars:\"\",yield_time_ms:30000,max_output_tokens:10000});\ntext(r.output);\n"
}

{
  "cell_id": "62",
  "yield_time_ms": 30000,
  "max_tokens": 10000
}

{
  "input": "const r = await tools.write_stdin({session_id:25923,chars:\"\",yield_time_ms:30000,max_output_tokens:10000});\ntext(r.output);\n"
}

{
  "cell_id": "63",
  "yield_time_ms": 30000,
  "max_tokens": 10000
}

{
  "input": "const r = await tools.exec_command({cmd:\"bash dev_eval.sh experiments/balexact.json && bash dev_eval.sh experiments/v2t085.json && bash dev_eval.sh experiments/v2t115.json\",\"workdir\":\"/workspace\",\"yield_time_ms\":1000,\"max_output_tokens\":12000});\ntext(r.output);\nif (r.session_id) text(`SESSION_ID=${r.session_id}`);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"cp train_nano.py train_diag.py\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":2000});\ntext(r.output);\n"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Update File: /workspace/train_diag.py\\n@@\\n mean_loss = sum_loss / max(1, nseen)\\n ppl = math.exp(mean_loss)\\n+quarter=[]\\n+for q in range(4):\\n+    qq=ho[q*250000:(q+1)*250000]\\n+    qwin=(len(qq)-1)//a.block; qsum=0.0; qseen=0\\n+    with torch.no_grad():\\n+        for i in range(0,qwin,a.batch):\\n+            bs=list(range(i,min(i+a.batch,qwin)))\\n+            xx=torch.stack([qq[s*a.block:s*a.block+a.block] for s in bs]).to(dev)\\n+            yy=torch.stack([qq[s*a.block+1:s*a.block+1+a.block] for s in bs]).to(dev)\\n+            with torch.autocast('cuda',dtype=torch.bfloat16): _,ll=model(xx,yy)\\n+            qsum+=ll.item()*len(bs); qseen+=len(bs)\\n+    quarter.append(math.exp(qsum/qseen))\\n+print('QUARTER_PPL',' '.join(f'{v:.3f}' for v in quarter))\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python - <<'PY'\\nimport json,pickle,numpy as np\\nz=pickle.load(open('scores.pkl','rb')); p=z['probs']; ids=z['ids']; idx={int(x):i for i,x in enumerate(ids)}; lens=np.load('token_lengths.npy')\\nfor f in ['balexact','v2eq','v2t085','v2t115']:\\n sel=json.load(open('experiments/'+f+'.json')); sm=np.zeros(4,int); n=np.zeros(4,int); tot=0\\n for d in sel:\\n  k=p[idx[d]].argmax(); sm[k]+=lens[d];n[k]+=1;tot+=lens[d]\\n  if tot>=12000000:break\\n print(f,tot,sm,n,n.sum())\\nPY\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":3000});\ntext(r.output);\n"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Add File: /workspace/page_fingerprints.py\\n+import json,hashlib,pickle,re\\n+pages=[]\\n+for line in open('/workspace/data/pool.jsonl'):\\n+    d=json.loads(line); sig=[]\\n+    for part in d['text'].split('<|endoftext|>'):\\n+        # Boundary-page fingerprints catch shifted crawl bundles.  Ignore tiny headers/fragments.\\n+        if len(part)>=400:\\n+            norm=' '.join(re.findall(r\\\"[a-z0-9]+\\\",part.lower()))\\n+            if len(norm)>=300:\\n+                h=int.from_bytes(hashlib.blake2b(norm.encode(),digest_size=8).digest(),'little')\\n+                sig.append((h,len(part)))\\n+    pages.append(sig)\\n+pickle.dump(pages,open('/workspace/pages.pkl','wb'),protocol=4)\\n+print('docs',len(pages),'pages',sum(map(len,pages)))\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\nconst r = await tools.exec_command({cmd:\"python page_fingerprints.py\",\"workdir\":\"/workspace\",\"yield_time_ms\":1000,\"max_output_tokens\":3000});\ntext(r.output);\nif (r.session_id) text(`SESSION_ID=${r.session_id}`);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:65549,chars:\"\",yield_time_ms:30000,max_output_tokens:12000});\ntext(r.output);\n"
}

{
  "cell_id": "69",
  "yield_time_ms": 30000,
  "max_tokens": 12000
}

{
  "input": "const r = await tools.write_stdin({session_id:65549,chars:\"\",yield_time_ms:30000,max_output_tokens:12000});\ntext(r.output);\n"
}

{
  "cell_id": "70",
  "yield_time_ms": 30000,
  "max_tokens": 12000
}