gpt-5.6-sol

{
  "cell_id": "98",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.write_stdin({\"session_id\":20110,\"chars\":\"\",\"yield_time_ms\":30000,\"max_output_tokens\":3000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "99",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.write_stdin({\"session_id\":20110,\"chars\":\"\",\"yield_time_ms\":30000,\"max_output_tokens\":3000}); text(JSON.stringify(r))\n"
}

{
  "input": "const r = await tools.exec_command({\"cmd\":\"python - <<'PY'\\nimport json\\nA=json.load(open('/tmp/hybridsel.json'));B=json.load(open('submission/selection.json'));print(A==B,sum(a==b for a,b in zip(A,B)),len(A),len(B));print([(i,A[i],B[i]) for i in range(len(A)) if A[i]!=B[i]][:20])\\nPY\\ncat /workspace/.timer/remaining_secs\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":3000}); text(r.output)\n"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Add File: /workspace/claim.md\\n+# Hypothesis\\n+\\n+Under a severe token budget, a balanced mixture of clean prose selected by both\\n+general quality and target-register similarity will train a better small language\\n+model than random web data or a single global quality ranking.  Fine target-style\\n+matching should help news and technical Q&A most; conservative semantic cues\\n+should transfer better for encyclopedic and general prose.\\n+\\n+# Mechanism\\n+\\n+The curator removes short, repetitive, non-English, menu-heavy, SEO-like, and\\n+template-heavy documents, suppresses normalized-prefix duplicates, and uses a\\n+source-quality prior observed at a corpus-wide acquisition boundary.  It learns\\n+smoothed one-vs-rest word log odds from the four equal disclosed dev quarters.\\n+News and technical Q&A use this target-aware ranker.  Encyclopedic and general\\n+web prose use a more conservative quality/semantic ranker, then are assigned to\\n+registers by the target classifier.  Token-weighted round-robin ordering prevents\\n+any register from being lost at the 12M-token cutoff.\\n+\\n+An observable prediction independent of final perplexity is that the produced\\n+prefix will have approximately equal estimated mass in all four registers (the\\n+report predicts about 4.0M tokens each before the deliberately overfilled tail),\\n+no duplicate IDs or normalized-prefix duplicates, and substantially fewer\\n+navigation/privacy/SEO phrases and repeated lines than a seeded random prefix.\\n+It should also classify into the disclosed news and technical quarters more\\n+accurately than the manual-only ranker.\\n+\\n+# Falsification\\n+\\n+The hypothesis is falsified if a seeded random 12M-token sample has no higher\\n+boilerplate/repetition rates, if any register is below 20% of the consumed prefix,\\n+or if ablating either quality filtering or register balancing does not worsen\\n+mean loss across repeated seeds.  It is also falsified mechanistically if the\\n+target-aware ranker fails to improve news/technical dev loss while the hybrid's\\n+gain comes solely from one domain.  These tests do not require the hidden score.\\n+\\n+# Transfer\\n+\\n+The quality, deduplication, and token-balanced mixture components should transfer\\n+to other English web pools and small autoregressive models.  The learned log-odds\\n+weights transfer only when the deployment mixture resembles the disclosed target;\\n+for a different language or domain mix, recompute them from a small representative\\n+target sample and change the quotas.  The acquisition-boundary prior is specific\\n+to this pool and should be re-estimated rather than copied elsewhere.\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\n"
}

{
  "input": "const r = await tools.exec_command({\"cmd\":\"chmod +x submission/curate.py\\npython -m py_compile submission/curate.py\\npython - <<'PY'\\nimport json,re,hashlib,os,numpy as np\\ns=json.load(open('submission/selection.json')); assert isinstance(s,list) and s and all(type(i) is int and 0<=i<182016 for i in s); assert len(s)==len(set(s))\\nwant=set(s); seen={}; dup=[]; chars=0\\nfor l in open('data/pool.jsonl'):\\n o=json.loads(l)\\n if o['id'] in want:\\n  chars+=len(o['text']); n=re.sub(r'\\\\W+',' ',o['text'][:16000].replace('<|endoftext|>',' ').lower()).strip(); h=hashlib.blake2b(n[:1800].encode(),digest_size=10).digest()\\n  if h in seen:dup.append((seen[h],o['id']))\\n  else:seen[h]=o['id']\\nprint({'ids':len(s),'unique':len(set(s)),'min':min(s),'max':max(s),'selected_chars':chars,'prefix_hash_duplicates':len(dup),'packed_tokens':len(np.load('/tmp/hybrid.npy'))})\\nprint('files',[(p,os.path.getsize(p)) for p in ['submission/selection.json','submission/curate.py','claim.md']])\\nPY\\nsha256sum submission/selection.json /tmp/hybridsel.json submission/curate.py claim.md\\ncat /workspace/.timer/remaining_secs\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":5000}); text(JSON.stringify(r))\n"
}

{
  "input": "const r = await tools.exec_command({\"cmd\":\"python - <<'PY'\\nimport sys,json\\nsys.path.insert(0,'submission');import curate\\ntw=curate.target_weights('data/multi_dev.npy');s=json.load(open('submission/selection.json'));want=set(s);info={}\\nfor l in open('data/pool.jsonl'):\\n o=json.loads(l)\\n if o['id'] in want:\\n  z=curate.features(o,tw);info[o['id']]={'d':max(z[1],key=z[1].get),'e':z[2]}\\njson.dump(info,open('/tmp/dominfo.json','w'))\\n# Preserve per-domain within-ranker order, change only token allocation.\\nby={d:[] for d in ('encyclopedic','web','news','technical')}\\nfor i in s:by[info[i]['d']].append(i)\\nfor label,w in [('web30',{'encyclopedic':.25,'web':.30,'news':.23,'technical':.22}),('enc30',{'encyclopedic':.30,'web':.25,'news':.23,'technical':.22})]:\\n idx={d:0 for d in by};used={d:0 for d in by};out=[];active=set(by);pr={'web':0,'news':1,'technical':2,'encyclopedic':3}\\n while active:\\n  d=min(active,key=lambda x:(used[x]/w[x],pr[x]))\\n  if idx[d]>=len(by[d]):active.remove(d);continue\\n  i=by[d][idx[d]];idx[d]+=1;out.append(i);used[d]+=info[i]['e']\\n json.dump(out,open('/tmp/'+label+'.json','w'));print(label,used)\\nPY\\npython pack_selection.py /tmp/web30.json /tmp/web30.npy\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":5000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "104",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({\"session_id\":93938,\"chars\":\"\",\"yield_time_ms\":30000,\"max_output_tokens\":5000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "105",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.exec_command({\"cmd\":\"PYTHONPATH=/workspace python /tmp/train_domains.py --train_npy /tmp/web30.npy --heldout_npy data/multi_dev.npy --out_json /tmp/web30.json.out --max_iters 3000 --seed 1337\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":3000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "106",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.write_stdin({\"session_id\":84673,\"chars\":\"\",\"yield_time_ms\":30000,\"max_output_tokens\":3000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "107",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.exec_command({\"cmd\":\"python pack_selection.py /tmp/enc30.json /tmp/enc30.npy\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":2000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "108",
  "yield_time_ms": 30000,
  "max_tokens": 2000
}

{
  "input": "const r = await tools.write_stdin({\"session_id\":23041,\"chars\":\"\",\"yield_time_ms\":30000,\"max_output_tokens\":2000}); text(JSON.stringify(r))\n"
}

{
  "input": "const r = await tools.exec_command({\"cmd\":\"PYTHONPATH=/workspace python /tmp/train_domains.py --train_npy /tmp/enc30.npy --heldout_npy data/multi_dev.npy --out_json /tmp/enc30.out --max_iters 3000 --seed 1337\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":3000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "110",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.write_stdin({\"session_id\":35891,\"chars\":\"\",\"yield_time_ms\":30000,\"max_output_tokens\":3000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "111",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}