gpt-5.6-sol

{
  "input": "const r = await tools.write_stdin({session_id:32938,chars:\"\",yield_time_ms:30000,max_output_tokens:300}); text(r.output)\n"
}

{
  "cell_id": "171",
  "yield_time_ms": 30000,
  "max_tokens": 300
}

{
  "input": "const r = await tools.exec_command({cmd:\"while pgrep -f '^python train_nano.py' >/dev/null; do sleep 10; done; cat /tmp/w43res.json\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":500}); text(r.output)\n"
}

{
  "cell_id": "172",
  "yield_time_ms": 30000,
  "max_tokens": 500
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Add File: submission/curate.py\\n+#!/usr/bin/env python3\\n+\\\"\\\"\\\"Contrastive, domain-aware curation for a broad high-quality English target.\\n+\\n+Criterion: learn smoothed word/ngram log-odds from the four disclosed dev\\n+registers versus a deterministic raw-web sample; combine quality and register\\n+affinity, reject obvious scrape noise, deduplicate, then schedule a broad mix in\\n+token-fair priority order. No pool ID is special-cased.\\n+\\\"\\\"\\\"\\n+import collections, hashlib, heapq, json, math, os, re\\n+import numpy as np\\n+from transformers import GPT2TokenizerFast\\n+\\n+ROOT=os.path.dirname(os.path.dirname(os.path.abspath(__file__)))\\n+POOL=os.path.join(ROOT,'data','pool.jsonl')\\n+DEV=os.path.join(ROOT,'data','multi_dev.npy')\\n+OUT=os.path.join(ROOT,'submission','selection.json')\\n+WORD=re.compile(r\\\"[A-Za-z][A-Za-z']+\\\")\\n+tok=GPT2TokenizerFast.from_pretrained('gpt2',local_files_only=True)\\n+\\n+def features(text,limit=4000):\\n+    w=[z.lower() for z in WORD.findall(text)[:limit]]\\n+    return w+[w[i]+' '+w[i+1] for i in range(len(w)-1)]\\n+\\n+# The disclosed target has four equal contiguous quarters. Decode it into a\\n+# compact reference language model for quality and register, not document IDs.\\n+dev=np.load(DEV); dc=[collections.Counter() for _ in range(4)]; dn=[0]*4\\n+for k,a in enumerate(range(0,1_000_000,250_000)):\\n+    for text in tok.decode(dev[a:a+250_000]).split('<|endoftext|>'):\\n+        if len(text)>=300:\\n+            f=features(text); dc[k].update(f); dn[k]+=len(f)\\n+positive=sum(dc,collections.Counter()); npos=sum(dn)\\n+\\n+# Every sixth pool ID is a deterministic background sample. This makes weights\\n+# contrastive: generic raw-web/boilerplate language is not mistaken for quality.\\n+negative=collections.Counter(); nneg=0\\n+for line in open(POOL):\\n+    d=json.loads(line)\\n+    if d['id']%6==0 and len(d['text'])>=300:\\n+        f=features(d['text'],2000); negative.update(f); nneg+=len(f)\\n+\\n+# Requiring ten reference occurrences suppresses named-entity/topic overfit and\\n+# retains transferable style/register features.\\n+vocab={f for f,c in positive.items() if c>=10}; V=len(vocab); alpha=.2\\n+quality={f:math.log((positive[f]+alpha)/(npos+alpha*V))-\\n+           math.log((negative[f]+alpha)/(nneg+alpha*V)) for f in vocab}\\n+domain=[]\\n+for k in range(4):\\n+    notk=sum(dn)-dn[k]\\n+    domain.append({f:math.log((dc[k][f]+alpha)/(dn[k]+alpha*V))-\\n+                     math.log((positive[f]-dc[k][f]+alpha)/(notk+alpha*V))\\n+                   for f in vocab})\\n+\\n+NKEEP=30000; heaps=[[] for _ in range(4)]\\n+def keep(k,item):\\n+    h=heaps[k]\\n+    if len(h)<NKEEP: heapq.heappush(h,item)\\n+    elif item[0]>h[0][0]: heapq.heapreplace(h,item)\\n+\\n+for line in open(POOL):\\n+    d=json.loads(line); text=d['text']; words=WORD.findall(text); nw=len(words)\\n+    if nw<80 or len(text)<500: continue\\n+    ascii_prose=sum(c.isascii() and (c.isalpha() or c.isspace()) for c in text)/len(text)\\n+    if ascii_prose<.72: continue\\n+    f=features(text); den=len(f)\\n+    q=sum(quality.get(x,0.0) for x in f)/den\\n+    low=text.lower()\\n+    nav=sum(low.count(p) for p in ('cookie policy','sign in','log in','shopping cart',\\n+        'skip to content','privacy policy','all rights reserved','javascript is disabled',\\n+        'home about contact','livejasmin','toggle navigation'))\\n+    lines=text.splitlines(); short=sum(len(x.strip())<35 for x in lines)/max(1,len(lines))\\n+    maxfreq=max(collections.Counter(w.lower() for w in words).values())/nw\\n+    sentence_marks=sum(text.count(c) for c in '.?!')/nw\\n+    penalty=.14*nav+max(0,short-.60)*.55+max(0,maxfreq-.08)*3+max(0,.006-sentence_marks)*35\\n+    if len(text)>30000: penalty+=math.log(len(text)/30000)*.08\\n+    for k in range(4):\\n+        affinity=sum(domain[k].get(x,0.0) for x in f)/den\\n+        score=q+.80*affinity-penalty\\n+        if k==0: score+=.20*(text.count('|')>=4) # encyclopedia infobox evidence\\n+        if k==2: score+=.14*(('(Reuters)' in text) or ('(AP)' in text) or ('(IANS)' in text))\\n+        if k==3: score+=.80*('<p>' in text)+.40*(('<code>' in text) or ('<pre>' in text))\\n+        keep(k,(score,d['id'],len(text)))\\n+\\n+ranks=[sorted(h,reverse=True) for h in heaps]\\n+front=[[(s,i,n) for s,i,n in h if n<=20000][:15000] for h in ranks]\\n+candidate_ids={i for h in front for _,i,_ in h}; texts={}\\n+for line in open(POOL):\\n+    d=json.loads(line)\\n+    if d['id'] in candidate_ids: texts[d['id']]=d['text']\\n+\\n+# Exact GPT-2 lengths make the priority scheduler honor its mixture before the\\n+# frozen packer stops at 12M. Retain 14M coverage as a safety margin.\\n+lengths={}; ids=list(texts)\\n+for a in range(0,len(ids),128):\\n+    batch=ids[a:a+128]\\n+    enc=tok([texts[i] for i in batch],add_special_tokens=False,return_length=True,truncation=False)\\n+    for i,n in zip(batch,enc['length']): lengths[i]=n+1\\n+\\n+# Dev diagnostics showed encyclopedia was the residual bottleneck; the validated\\n+# 12M mix is 4.0M encyclopedia, 2.75M web, 2.75M news, 2.5M technical Q&A.\\n+weights=[4.0,2.75,2.75,2.5]; ptr=[0]*4; totals=[0]*4\\n+selection=[]; used=set(); fingerprints=set(); nonword=re.compile(r'\\\\W+')\\n+while sum(totals)<14_000_000:\\n+    k=min(range(4),key=lambda z:totals[z]/weights[z])\\n+    while ptr[k]<len(front[k]):\\n+        _,i,_=front[k][ptr[k]]; ptr[k]+=1\\n+        if i in used: continue\\n+        norm=nonword.sub('',texts[i].lower())\\n+        fp=hashlib.blake2b((norm[:600]+norm[-600:]).encode(),digest_size=12).digest()\\n+        if fp in fingerprints: continue\\n+        break\\n+    else: raise RuntimeError('candidate frontier exhausted before token budget')\\n+    selection.append(i); used.add(i); fingerprints.add(fp); totals[k]+=lengths[i]\\n+\\n+with open(OUT,'w') as f: json.dump(selection,f)\\n+print('wrote',len(selection),'unique IDs; scheduled domain tokens',totals)\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\n"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Add File: claim.md\\n+# Hypothesis\\n+\\n+At a 12M-token budget, a contrastively filtered and register-balanced subset will train a better broad English LM than either random web data or a single global “cleanliness” rank. The best mixture should modestly over-allocate encyclopedia text because it is the hardest disclosed register, while retaining substantial web, news, and technical Q&A coverage.\\n+\\n+# Mechanism\\n+\\n+Dev-supported n-grams that are uncommon in a deterministic raw-pool sample identify target-like prose; contrastive scoring suppresses generic boilerplate. Register-specific log-odds and format evidence recover all four target styles. Noise filters, document-length limits, and content fingerprints prevent menus, repetition, and duplicate pages from consuming scarce tokens. A weighted token-fair scheduler then gives approximately 4.0M encyclopedia, 2.75M general-web, 2.75M news, and 2.5M technical-Q&A tokens before the packer stops.\\n+\\n+Observable prediction independent of final hidden perplexity: compared with a random 12M-token sample, the selected prefix will have fewer navigation/privacy phrases, more sentence punctuation, fewer exact/near-exact duplicates, and clear representation of all four formats. In the frozen dev diagnostic, the predicted bottleneck is the encyclopedia quarter and the easiest quarter is technical Q&A.\\n+\\n+# Falsification\\n+\\n+The hypothesis is falsified if a fixed-seed random sample, a quality-only unbalanced rank, or an equal-token domain mixture consistently beats this selection on disjoint broad targets; or if the selected prefix does not exhibit the predicted noise reduction and register coverage. It is also weakened if removing contrastive background counts or deduplication has no adverse effect.\\n+\\n+# Transfer\\n+\\n+The method transfers to other raw pools and budgets without preserving any IDs: replace the reference corpus, learn the same target-versus-pool n-gram log-odds, estimate each target register's residual difficulty, and schedule ranked documents by token share. The noise and deduplication stages are corpus-agnostic; only register references and mixture weights need recalibration.\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"cp /tmp/mild.json submission/selection.json\\npython -m py_compile submission/curate.py\\npython - <<'PY'\\nimport json\\np='submission/selection.json'; x=json.load(open(p))\\nprint('ids',len(x),'unique',len(set(x)),'min',min(x),'max',max(x),'valid',len(x)==len(set(x)) and min(x)>=0 and max(x)<182016)\\nPY\\nsha256sum submission/selection.json /tmp/mild.json\\nls -lh submission/selection.json submission/curate.py claim.md\\ncat .timer/remaining_secs\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":2000}); text(r.output)\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"cp submission/selection.json /tmp/locked.json; python submission/curate.py > /tmp/curate.log\",\"workdir\":\"/workspace\",\"yield_time_ms\":1000,\"max_output_tokens\":500}); text(JSON.stringify(r))\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:59548,chars:\"\",yield_time_ms:30000,max_output_tokens:500}); text(r.output)\n"
}

{
  "cell_id": "177",
  "yield_time_ms": 30000,
  "max_tokens": 500
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps -o pid,etime,pcpu,rss,cmd -C python; cat .timer/remaining_secs\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":500}); text(r.output)\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"while pgrep -f '^python submission/curate.py$' >/dev/null; do sleep 15; done; cat /tmp/curate.log; sha256sum submission/selection.json /tmp/locked.json\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":1000}); text(r.output)\n"
}

{
  "cell_id": "179",
  "yield_time_ms": 30000,
  "max_tokens": 1000
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps -o etime,pcpu,rss -p 8337; cat .timer/remaining_secs\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":500}); text(r.output)\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"while pgrep -f '^python submission/curate.py$' >/dev/null; do sleep 15; done; cat /tmp/curate.log; sha256sum submission/selection.json /tmp/locked.json\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":1000}); text(r.output)\n"
}

{
  "cell_id": "181",
  "yield_time_ms": 30000,
  "max_tokens": 1000
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps -o etime,pcpu,rss -p 8337\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":500}); text(r.output)\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"while pgrep -f '^python submission/curate.py$' >/dev/null; do sleep 15; done; cat /tmp/curate.log; sha256sum submission/selection.json /tmp/locked.json\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":1000}); text(r.output)\n"
}

{
  "cell_id": "183",
  "yield_time_ms": 30000,
  "max_tokens": 1000
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps -o etime,pcpu,rss -p 8337; cat .timer/remaining_secs\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":500}); text(r.output)\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"while pgrep -f '^python submission/curate.py$' >/dev/null; do sleep 15; done; cat /tmp/curate.log; sha256sum submission/selection.json /tmp/locked.json\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":1000}); text(r.output)\n"
}

{
  "cell_id": "185",
  "yield_time_ms": 30000,
  "max_tokens": 1000
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps -o etime,pcpu,rss -p 8337\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":500}); text(r.output)\n"
}