gpt-5.6-sol

{
  "input": "const r = await tools.exec_command({cmd:\"ps -o pid,etime,pcpu,rss -p 1263; ls -lh data/word_scores.npy 2>/dev/null || true; cat .timer/remaining_secs\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":1000}); text(r.output);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:43105,chars:\"\",yield_time_ms:30000,max_output_tokens:1000}); text(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps -o pid,etime,pcpu,rss -p 1263; ls -lh data/word_scores.npy 2>/dev/null || true\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":1000}); text(r.output);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:43105,chars:\"\",yield_time_ms:30000,max_output_tokens:1000}); text(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps -o pid,etime,pcpu,rss -p 1263; ls -lh data/word_scores.npy 2>/dev/null || true\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":1000}); text(r.output);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:43105,chars:\"\",yield_time_ms:30000,max_output_tokens:1000}); text(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps -o pid,etime,pcpu,rss -p 1263; ls -lh data/word_scores.npy 2>/dev/null || true\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":1000}); text(r.output);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:43105,chars:\"\",yield_time_ms:30000,max_output_tokens:1000}); text(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps -o pid,etime,pcpu,rss -p 1263; ls -lh data/word_scores.npy 2>/dev/null || true\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":1000}); text(r.output);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:43105,chars:\"\",yield_time_ms:30000,max_output_tokens:1000}); text(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps -o pid,etime,pcpu,rss -p 1263; ls -lh data/word_scores.npy 2>/dev/null || true\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":1000}); text(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python3 - <<'PY'\\nimport json,random\\nx=list(range(182016));random.Random(42).shuffle(x);json.dump(x,open('/tmp/random.json','w'))\\nPY\\nbash dev_eval.sh /tmp/random.json\",\"workdir\":\"/workspace\",\"yield_time_ms\":1000,\"max_output_tokens\":2000}); text(JSON.stringify(r));\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:52602,chars:\"\",yield_time_ms:30000,max_output_tokens:2000}); text(r.output);\n"
}

{
  "cell_id": "55",
  "yield_time_ms": 30000,
  "max_tokens": 2000
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps -o pid,etime,pcpu,rss,cmd -C python3; ls -lh /tmp/dev_train.npy /tmp/dev.json 2>/dev/null || true\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":2000}); text(r.output);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:52602,chars:\"\",yield_time_ms:30000,max_output_tokens:2000}); text(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps -o pid,etime,pcpu,rss,cmd -p 1263,1602; ls -lh /tmp/dev_train.npy 2>/dev/null || true\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":1000}); text(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps aux | rg 'train_nano|word_score|dev_eval' | head; ls -lh data/word_scores.npy 2>/dev/null || true\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":1000}); text(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python3 - <<'PY'\\nimport json,numpy as np,re\\ns=np.load('data/word_scores.npy'); texts={}\\nfor l in open('data/pool.jsonl'):\\n d=json.loads(l);texts[d['id']]=d['text']\\nfor k,name in enumerate(['wiki','web','news','qa'],1):\\n print('\\\\n##',name)\\n for row in s[np.argsort(s[:,k])[-10:][::-1]]:\\n  i=int(row[0]); print(i,round(row[k],3),len(texts[i]),re.sub(r'\\\\s+',' ',texts[i])[:150])\\nprint('quantiles',np.quantile(s[:,1:], [.1,.5,.9,.99],axis=0))\\nPY\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":10000}); text(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"ps -o pid,etime,pcpu,rss,cmd -C python3 | tail\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":1000}); text(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python3 - <<'PY'\\nimport json,numpy as np,re\\ns=np.load('data/word_scores.npy'); tops=[]\\nfor k in range(1,5): tops.append([int(x) for x in s[np.argsort(s[:,k])[-10:][::-1],0]])\\nwant=set(sum(tops,[]));texts={}\\nfor l in open('data/pool.jsonl'):\\n d=json.loads(l)\\n if d['id'] in want:texts[d['id']]=d['text']\\nfor k,name in enumerate(['wiki','web','news','qa']):\\n print('\\\\n##',name)\\n for i in tops[k]:print(i,round(float(s[i,k+1]),3),len(texts[i]),re.sub(r'\\\\s+',' ',texts[i])[:150])\\nprint('quantiles',np.quantile(s[:,1:], [.1,.5,.9,.99],axis=0))\\nPY\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":12000}); text(r.output);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:52602,chars:\"\",yield_time_ms:30000,max_output_tokens:2000}); text(r.output);\n"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Add File: /workspace/submission/curate.py\\n+#!/usr/bin/env python3\\n+\\\"\\\"\\\"Rank clean documents by quality and fit to four disclosed target registers.\\n+\\n+The target's four equal token quarters provide vocabulary distributions for\\n+encyclopedic, edited-web, news, and technical-Q&A registers.  These are used as\\n+weak log-odds features, not as a source of training text.  Structural filters\\n+remove boilerplate, non-English, fragments, repetition, and SEO/navigation\\n+pages.  A token-weighted fair merge keeps all four registers represented.\\n+\\\"\\\"\\\"\\n+import collections, json, math, os, re\\n+import numpy as np\\n+from transformers import AutoTokenizer\\n+\\n+ROOT = \\\"/workspace\\\"\\n+POOL = ROOT + \\\"/data/pool.jsonl\\\"\\n+DEV = ROOT + \\\"/data/multi_dev.npy\\\"\\n+OUT = ROOT + \\\"/submission/selection.json\\\"\\n+WORD = re.compile(r\\\"[A-Za-z]+(?:'[A-Za-z]+)?\\\", re.ASCII)\\n+STOP = set(\\\"the a an and or but if while of to in on for with by from at as is are was were be been being this that these those it its they their he she we you not no can could would should may will have has had do does did than then also into about over after before between through during such other more most some any each which who what when where how\\\".split())\\n+BAD = (\\\"cookie policy\\\", \\\"privacy policy\\\", \\\"all rights reserved\\\", \\\"skip to content\\\",\\n+       \\\"shopping cart\\\", \\\"add to cart\\\", \\\"log in\\\", \\\"sign up\\\", \\\"sitemap\\\",\\n+       \\\"toggle navigation\\\", \\\"javascript is required\\\", \\\"search results for\\\",\\n+       \\\"wordpress theme\\\", \\\"casino\\\", \\\"viagra\\\", \\\"porn\\\", \\\"bestiality\\\")\\n+TECH = set(\\\"code function class method error file data python java javascript linux windows server database api algorithm variable object array string command install software application network system using return output input\\\".split())\\n+NEWS = set(\\\"said says told reported according officials government president minister police court company percent year monday tuesday wednesday thursday friday saturday sunday\\\".split())\\n+\\n+def words(text):\\n+    return WORD.findall(text.lower())\\n+\\n+def main():\\n+    tokenizer = AutoTokenizer.from_pretrained(\\\"gpt2\\\")\\n+    dev = np.load(DEV)\\n+    dcounts=[]\\n+    for k in range(4):\\n+        dcounts.append(collections.Counter(words(tokenizer.decode(dev[k*250000:(k+1)*250000]))))\\n+    # Only background-count target vocabulary; this is much smaller than the\\n+    # raw web vocabulary and is all that the log-odds calculation needs.\\n+    vocab={w for c in dcounts for w,n in c.items() if n >= 3}\\n+    bg=collections.Counter(); bg_total=0\\n+    for line in open(POOL, encoding=\\\"utf-8\\\"):\\n+        text=json.loads(line)[\\\"text\\\"][:16000]\\n+        ws=words(text); bg_total += len(ws)\\n+        bg.update(w for w in ws if w in vocab)\\n+    dtotal=[sum(c.values()) for c in dcounts]\\n+    # A conservative denominator vocabulary makes the feature about relative\\n+    # register usage rather than rewarding words absent from the pool.\\n+    V=max(50000,len(vocab))\\n+    ratios=[]\\n+    for c,nt in zip(dcounts,dtotal):\\n+        rr={}\\n+        for w,n in c.items():\\n+            if n >= 3:\\n+                z=math.log((n+1)/(nt+V))-math.log((bg.get(w,0)+1)/(bg_total+V))\\n+                rr[w]=max(-2.0,min(2.0,z))\\n+        ratios.append(rr)\\n+\\n+    rows=[]\\n+    for line in open(POOL, encoding=\\\"utf-8\\\"):\\n+        d=json.loads(line); text=d[\\\"text\\\"]; low=text.lower(); ws=words(text[:30000]); n=len(ws)\\n+        if n < 180 or len(text) < 900 or len(text) > 80000:\\n+            continue\\n+        chars=max(1,len(text)); alpha=sum(ch.isalpha() for ch in text)/chars\\n+        ascii_ratio=sum(ord(ch)<128 for ch in text)/chars\\n+        stop=sum(w in STOP for w in ws)/n\\n+        sent=len(re.findall(r\\\"[.!?](?:[\\\\\\\"')\\\\]]|\\\\s|$)\\\",text))/n\\n+        urls=len(re.findall(r\\\"https?://|www\\\\.\\\",low))/n\\n+        bad=sum(low.count(x) for x in BAD)\\n+        lines=[re.sub(r\\\"\\\\s+\\\",\\\" \\\",x.strip().lower()) for x in text.splitlines() if len(x.strip())>30]\\n+        rep=0.0 if not lines else 1-len(set(lines))/len(lines)\\n+        # Smooth English prose quality, centered on edited target statistics.\\n+        q=2.0*min(1.0,n/500.0)\\n+        q-=5.0*abs(stop-0.39)\\n+        q-=4.0*max(0.0,0.66-alpha)\\n+        q-=5.0*max(0.0,0.91-ascii_ratio)\\n+        q-=18.0*max(0.0,0.018-sent)\\n+        q-=8.0*max(0.0,sent-0.11)\\n+        q-=min(3.0,35.0*urls) + min(4.0,0.35*bad) + 3.0*rep\\n+        if alpha < .52 or ascii_ratio < .78 or stop < .20 or sent < .006 or bad >= 8 or rep > .35:\\n+            continue\\n+        sample=ws[:3000]; sums=[0.0]*4; cnt=[0]*4\\n+        for w in sample:\\n+            for k,rr in enumerate(ratios):\\n+                z=rr.get(w)\\n+                if z is not None: sums[k]+=z; cnt[k]+=1\\n+        rel=[s/max(1,c) for s,c in zip(sums,cnt)]\\n+        wset=set(sample)\\n+        # Small, transparent genre cues counter topic noise in unigram log odds.\\n+        formal=(low.count(\\\" according to \\\")+low.count(\\\" was born \\\")+low.count(\\\" is a \\\"))/max(1,n/500)\\n+        news=len(wset & NEWS)/len(NEWS)\\n+        tech=len(wset & TECH)/len(TECH) + min(.12,text.count(\\\"?\\\")/max(1,n))\\n+        bonus=[min(.25,.025*formal),0.0,1.4*news,1.8*tech]\\n+        scores=[q + 2.2*r + b for r,b in zip(rel,bonus)]\\n+        # GPT-2 tokens are close to chars/4 for English; this estimate is used\\n+        # only for mixture balancing, never for validity or final truncation.\\n+        est=max(1,int(chars/4.0)+1)\\n+        rows.append((d[\\\"id\\\"],est,scores))\\n+\\n+    # Put each document in its best-matching register after centering scores;\\n+    # this avoids duplicates while retaining four genuinely distinct queues.\\n+    mat=np.asarray([r[2] for r in rows])\\n+    center=np.median(mat,axis=0); scale=np.maximum(.05,np.std(mat,axis=0))\\n+    queues=[[] for _ in range(4)]\\n+    for (did,est,sc),zz in zip(rows,(mat-center)/scale):\\n+        k=int(np.argmax(zz))\\n+        queues[k].append((sc[k],did,est))\\n+    for q in queues:q.sort(reverse=True)\\n+    pos=[0]*4; used_tokens=[0]*4; order=[]\\n+    # Fair scheduling by estimated tokens, with extra tail coverage for safety.\\n+    target=18_000_000\\n+    while sum(used_tokens)<target:\\n+        available=[k for k in range(4) if pos[k]<len(queues[k])]\\n+        if not available: break\\n+        k=min(available,key=lambda j:used_tokens[j])\\n+        _,did,est=queues[k][pos[k]];pos[k]+=1\\n+        order.append(did);used_tokens[k]+=est\\n+    # Append all remaining valid documents in global score order, ensuring the\\n+    # official packer can always fill even if the estimate is unusually low.\\n+    chosen=set(order)\\n+    tail=sorted(((max(sc),did) for did,_,sc in rows if did not in chosen),reverse=True)\\n+    order.extend(did for _,did in tail)\\n+    os.makedirs(os.path.dirname(OUT),exist_ok=True)\\n+    with open(OUT,\\\"w\\\") as f: json.dump(order,f)\\n+    print(\\\"selected\\\",len(order),\\\"priority ids; balanced estimates\\\",used_tokens)\\n+\\n+if __name__ == \\\"__main__\\\": main()\\n*** End Patch\"; text(await tools.apply_patch(patch));\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python3 submission/curate.py\",\"workdir\":\"/workspace\",\"yield_time_ms\":1000,\"max_output_tokens\":1000}); text(JSON.stringify(r));\n"
}