gpt-5.6-sol

{
  "input": "const r = await tools.exec_command({cmd:\"python - <<'PY'\\nimport pickle,json,numpy as np\\np=pickle.load(open('scores.pkl','rb'))['probs'];le=np.load('token_lengths.npy');e=np.load('english_ratio.npy');s=json.load(open('experiments/v3w160t130.json'));tot=0;u=[]\\nfor d in s:u.append(d);tot+=le[d];\\n# correct prefix\\nss=[];tot=0\\nfor d in s:ss.append(d);tot+=le[d];\\n# again break\\nss=[];tot=0\\nfor d in s:\\n ss.append(d);tot+=le[d]\\n if tot>=12e6:break\\nss=np.array(ss)\\nfor k in range(4):\\n x=ss[p[ss].argmax(1)==k]; print(k,len(x),np.quantile(e[x],[0,.01,.05,.1,.5]),'tok<.12',le[x][e[x]<.12].sum(),'docs',sum(e[x]<.12))\\nPY\",\"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 pickle,json,numpy as np,random\\np=pickle.load(open('scores.pkl','rb'))['probs'];le=np.load('token_lengths.npy');e=np.load('english_ratio.npy');s=json.load(open('experiments/v3w160t130.json'));u=[];tot=0\\nfor d in s:\\n if p[d].argmax()==3 and e[d]<.12:u.append(d)\\n tot+=le[d]\\n if tot>=12e6:break\\nrandom.seed(3);want=set(random.sample(u,min(25,len(u))))\\nfor l in open('data/pool.jsonl'):\\n d=json.loads(l)\\n if d['id'] in want:print('\\\\n',d['id'],e[d['id']],le[d['id']],repr(d['text'][:350]))\\nPY\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":12000});\ntext(r.output);\n"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Update File: /workspace/make_exp.py\\n@@\\n     ex=pickle.load(open('/workspace/extra.pkl','rb'))\\n+    english=np.load('/workspace/english_ratio.npy')\\n@@\\n     qual3=qual2-np.where(valid,0,5.0)\\n+    english_ok=(english>=.12)|((english>=.07)&(tech>=5))\\n+    qual6=qual3-np.where(english_ok,0,5.0)\\n@@\\n-        if mode.startswith(('v2','v3','v4','v5')):\\n-            qual=qual3 if mode.startswith(('v3','v4','v5')) else qual2\\n+        if mode.startswith(('v2','v3','v4','v5','v6')):\\n+            qual=qual6 if mode.startswith('v6') else (qual3 if mode.startswith(('v3','v4','v5')) else qual2)\\n@@\\n-            if mode.startswith(('v2','v3','v4','v5')):\\n+            if mode.startswith(('v2','v3','v4','v5','v6')):\\n@@\\n         if mode=='v5_w160t130': weights=[1.60,1,1,1.30]\\n+        if mode=='v6_w160t130': weights=[1.60,1,1,1.30]\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\nconst r = await tools.exec_command({cmd:\"python make_exp.py v6_w160t130 experiments/v6w160t130.json\",\"workdir\":\"/workspace\",\"yield_time_ms\":1000,\"max_output_tokens\":5000});\ntext(r.output);\nif (r.session_id) text(`SESSION_ID=${r.session_id}`);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:54630,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r.output);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:35720,chars:\"\",yield_time_ms:30000,max_output_tokens:10000});\ntext(r.output);\n"
}

{
  "cell_id": "147",
  "yield_time_ms": 30000,
  "max_tokens": 10000
}

{
  "input": "const r = await tools.write_stdin({session_id:35720,chars:\"\",yield_time_ms:30000,max_output_tokens:10000});\ntext(r.output);\n"
}

{
  "cell_id": "148",
  "yield_time_ms": 30000,
  "max_tokens": 10000
}

{
  "input": "const r = await tools.write_stdin({session_id:35720,chars:\"\",yield_time_ms:30000,max_output_tokens:10000});\ntext(r.output);\n"
}

{
  "cell_id": "149",
  "yield_time_ms": 30000,
  "max_tokens": 10000
}

{
  "input": "const r = await tools.write_stdin({session_id:35720,chars:\"\",yield_time_ms:30000,max_output_tokens:10000});\ntext(r.output);\n"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Update File: /workspace/make_exp.py\\n@@\\n         if mode=='v6_w160t130': weights=[1.60,1,1,1.30]\\n+        if mode=='v6_w160t140': weights=[1.60,1,1,1.40]\\n+        if mode=='v5_w160t140': weights=[1.60,1,1,1.40]\\n+        if mode=='v3_w160t150': weights=[1.60,1,1,1.50]\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\nconst r = await tools.exec_command({cmd:\"python make_exp.py v6_w160t140 experiments/v6w160t140.json && python make_exp.py v5_w160t140 experiments/v5w160t140.json && python make_exp.py v3_w160t150 experiments/v3w160t150.json\",\"workdir\":\"/workspace\",\"yield_time_ms\":1000,\"max_output_tokens\":5000});\ntext(r.output);\nif (r.session_id) text(`SESSION_ID=${r.session_id}`);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:43644,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"bash dev_eval.sh experiments/v6w160t140.json && bash dev_eval.sh experiments/v5w160t140.json && bash dev_eval.sh experiments/v3w160t150.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 patch = \"*** Begin Patch\\n*** Add File: /workspace/submission/curate.py\\n+#!/usr/bin/env python3\\n+\\\"\\\"\\\"Target-density curation with quality gates, exact deduplication, and domain balancing.\\n+\\n+The disclosed dev set supplies *distributional*, not document-ID, supervision.  A\\n+smoothed word/bigram density ratio ranks pool prose against a deterministic random\\n+pool reservoir.  Four target-quarter Naive Bayes heads provide register affinity.\\n+The final priority stream is a weighted-fair merge (encyclopedic 1.60, general web\\n+1.00, news 1.00, technical Q&A 1.40) measured in exact GPT-2 tokens.  Readability,\\n+length, boilerplate and corruption gates demote pathological raw pages.  Exact text\\n+hashes are emitted once.  No pool ID is named or treated specially.\\n+\\\"\\\"\\\"\\n+import argparse, hashlib, json, math, random, re\\n+from collections import Counter\\n+from pathlib import Path\\n+\\n+import numpy as np\\n+from transformers import GPT2TokenizerFast\\n+\\n+WORD_RE = re.compile(r\\\"[a-z][a-z0-9_'-]+\\\")\\n+PLAIN_WORD_RE = re.compile(r\\\"[A-Za-z]+\\\")\\n+TECH_RE = re.compile(r\\\"\\\\b(?:python|javascript|java|php|linux|ubuntu|android|jquery|sql|database|array|function|class|method|variable|server|client|api|compiler|programming|code|error|exception|command|script|html|css|xml|github|software|install|application)\\\\b\\\", re.I)\\n+QA_RE = re.compile(r\\\"(?:\\\\bhow (?:do|can|would|to)\\\\b|\\\\bi (?:am|have|'m) (?:trying|using|working)|share\\\\|improve|up vote|down vote|add a comment|asked [A-Z][a-z]{2} |stack (?:overflow|exchange))\\\", re.I)\\n+NEWS_RE = re.compile(r\\\"\\\\b(?:Reuters|Associated Press|AFP|AP\\\\)|said (?:on |in |at )|reported|spokesman|officials?|minister|government|president|police|court|according to)\\\\b\\\", re.I)\\n+WIKI_RE = re.compile(r\\\"(?:from Wikipedia|free encyclopedia|\\\\|Born\\\\|\\\\||\\\\|Died\\\\|\\\\||\\\\|Preceded by\\\\|\\\\||\\\\bReferences\\\\b|\\\\bExternal links\\\\b|\\\\[[0-9]+\\\\])\\\", re.I)\\n+BAD_RE = re.compile(r\\\"(?:cookie policy|sign in|log in|register|skip to (?:main )?content|shopping cart|all rights reserved|privacy policy|terms (?:and|of) (?:conditions|use)|subscribe|toggle navigation|search for:|wordpress themes)\\\", re.I)\\n+COMMON = set(\\\"the of and to in a is that for it on with as are was be by this from or at an have not but which you has can will they their if one all we more about would when who what there been were do no into than other also its may these your how our some such only out up my so them he she his her\\\".split())\\n+\\n+\\n+def terms(text):\\n+    w = WORD_RE.findall(text.lower())\\n+    return w + [w[i] + \\\" \\\" + w[i + 1] for i in range(len(w) - 1)]\\n+\\n+\\n+def prose_stats(t):\\n+    n = max(len(t), 1)\\n+    words = re.findall(r\\\"[A-Za-z]+(?:'[A-Za-z]+)?\\\", t)\\n+    nw = max(len(words), 1)\\n+    lines = t.splitlines() or [\\\"\\\"]\\n+    alpha = sum(c.isalpha() for c in t) / n\\n+    printable = sum(c.isprintable() or c in \\\"\\\\n\\\\t\\\" for c in t) / n\\n+    avg = sum(map(len, words)) / nw\\n+    sent = sum(t.count(x) for x in \\\".?!\\\") / nw\\n+    upper = sum(w.isupper() and len(w) > 1 for w in words) / nw\\n+    nav = sum(t.lower().count(x) for x in [\\\" login \\\", \\\" sign up\\\", \\\"cookie\\\", \\\"privacy policy\\\", \\\"skip to content\\\", \\\"all rights reserved\\\", \\\"contact us\\\", \\\"terms of use\\\", \\\"shopping cart\\\"])\\n+    repeat = max(Counter(x.strip().lower() for x in lines if len(x.strip()) > 20).values(), default=1) / max(len(lines), 1)\\n+    return alpha, printable, avg, sent, upper, nav / max(nw / 100, 1), repeat, nw\\n+\\n+\\n+def main():\\n+    ap = argparse.ArgumentParser()\\n+    ap.add_argument(\\\"--pool\\\", default=\\\"/workspace/data/pool.jsonl\\\")\\n+    ap.add_argument(\\\"--target\\\", default=\\\"/workspace/data/multi_dev.npy\\\")\\n+    ap.add_argument(\\\"--output\\\", default=\\\"/workspace/submission/selection.json\\\")\\n+    args = ap.parse_args()\\n+    tok = GPT2TokenizerFast.from_pretrained(\\\"gpt2\\\", local_files_only=True)\\n+\\n+    # The benchmark discloses four equal, ordered target registers.\\n+    target = np.load(args.target)\\n+    domains = []\\n+    for k in range(4):\\n+        decoded = tok.decode(target[k * 250_000:(k + 1) * 250_000])\\n+        docs = [z.strip() for z in decoded.split(\\\"<|endoftext|>\\\") if len(z.strip()) > 200]\\n+        chunks = []\\n+        for doc in docs:\\n+            for i in range(0, len(doc), 3500):\\n+                z = doc[i:i + 5000]\\n+                if len(z) > 500:\\n+                    chunks.append(z)\\n+        domains.append(chunks)\\n+\\n+    # Deterministic pool background for the density ratio.\\n+    rng = random.Random(1337)\\n+    reservoir = []\\n+    with open(args.pool) as f:\\n+        for i, line in enumerate(f):\\n+            t = json.loads(line)[\\\"text\\\"]\\n+            if len(t) > 300:\\n+                if len(reservoir) < 12_000:\\n+                    reservoir.append(t[:6000])\\n+                else:\\n+                    j = rng.randrange(i + 1)\\n+                    if j < len(reservoir):\\n+                        reservoir[j] = t[:6000]\\n+\\n+    pc, nc = Counter(), Counter()\\n+    dc = [Counter() for _ in range(4)]\\n+    for k, ds in enumerate(domains):\\n+        for text in ds:\\n+            tt = terms(text)\\n+            dc[k].update(tt)\\n+            pc.update(tt)\\n+    for text in reservoir:\\n+        nc.update(terms(text))\\n+    vocab = {z for z, n in pc.items() if n >= 3 and n + nc[z] >= 5}\\n+    V = len(vocab)\\n+    pt = sum(pc[z] for z in vocab)\\n+    nt = sum(nc[z] for z in vocab)\\n+    ratio = {z: math.log((pc[z] + 2) / (pt + 2 * V)) - math.log((nc[z] + 2) / (nt + 2 * V)) for z in vocab}\\n+    dt = [sum(c[z] for z in vocab) for c in dc]\\n+    dlog = [{z: math.log((dc[k][z] + 1) / (dt[k] + V)) for z in vocab} for k in range(4)]\\n+    del pc, nc, dc, reservoir, domains\\n+\\n+    ids, texts, qscore, probs, rawstats, extras, hashes, english = [], [], [], [], [], [], [], []\\n+\\n+    def flush_scores():\\n+        for s in texts:\\n+            tt = terms(s)\\n+            vv = [z for z in tt if z in vocab]\\n+            den = max(len(vv), 1)\\n+            qscore.append(sum(ratio[z] for z in vv) / den)\\n+            ll = np.array([sum(dlog[k][z] for z in vv) / den for k in range(4)])\\n+            ll = (ll - ll.max()) * 12\\n+            pp = np.exp(ll)\\n+            probs.append(pp / pp.sum())\\n+        texts.clear()\\n+\\n+    with open(args.pool) as f:\\n+        for line in f:\\n+            d = json.loads(line)\\n+            did, t = d[\\\"id\\\"], d[\\\"text\\\"]\\n+            ids.append(did)\\n+            texts.append(t[:20_000])\\n+            rawstats.append(prose_stats(t))\\n+            w = PLAIN_WORD_RE.findall(t)\\n+            nw = max(len(w), 1)\\n+            tech = len(TECH_RE.findall(t)) * 100 / nw + len(QA_RE.findall(t)) * 180 / nw\\n+            html = (t.count(\\\"<p>\\\") + 2 * t.count(\\\"<code>\\\") + 2 * t.count(\\\"<pre>\\\") + .2 * t.count(\\\"&quot;\\\") + .1 * t.count(\\\"&gt;\\\")) * 100 / nw\\n+            news = len(NEWS_RE.findall(t)) * 100 / nw\\n+            wiki = len(WIKI_RE.findall(t)) * 100 / nw\\n+            bad = len(BAD_RE.findall(t)) * 100 / nw\\n+            codechars = sum(t.count(x) for x in [\\\"{\\\", \\\"}\\\", \\\"();\\\", \\\"</\\\", \\\"/>\\\", \\\"=&gt;\\\", \\\"::\\\", \\\"()\\\"])\\n+            weird = sum((not c.isprintable()) and c not in \\\"\\\\n\\\\t\\\" for c in t) / max(len(t), 1)\\n+            moj = t.count(\\\"\\\") + t.count(\\\"�\\\") + t.count(\\\"Ã\\\") + t.count(\\\"Â\\\")\\n+            extras.append((len(t), tech, html, news, wiki, bad, codechars * 100 / nw, weird, moj))\\n+            hashes.append(int.from_bytes(hashlib.blake2b(t.encode(), digest_size=8).digest(), \\\"little\\\"))\\n+            wl = [x.lower() for x in w]\\n+            english.append(sum(x in COMMON for x in wl) / nw)\\n+            if len(texts) >= 2048:\\n+                flush_scores()\\n+    flush_scores()\\n+\\n+    ids = np.asarray(ids, dtype=np.int64)\\n+    qscore = np.asarray(qscore)\\n+    probs = np.asarray(probs)\\n+    s = np.asarray(rawstats)\\n+    ex = np.asarray(extras)\\n+    hashes = np.asarray(hashes, dtype=np.uint64)\\n+    english = np.asarray(english)\\n+\\n+    # Exact GPT-2 costs make every prefix obey the intended mixture.\\n+    lengths = np.zeros(int(ids.max()) + 1, dtype=np.int32)\\n+    batch, bids = [], []\\n+\\n+    def flush_lengths():\\n+        if not batch:\\n+            return\\n+        enc = tok(batch, add_special_tokens=False, return_length=True, truncation=False)\\n+        for did, n in zip(bids, enc[\\\"length\\\"]):\\n+            lengths[did] = n + 1\\n+        batch.clear(); bids.clear()\\n+\\n+    with open(args.pool) as f:\\n+        for line in f:\\n+            d = json.loads(line)\\n+            batch.append(d[\\\"text\\\"]); bids.append(d[\\\"id\\\"])\\n+            if len(batch) >= 512:\\n+                flush_lengths()\\n+    flush_lengths()\\n+\\n+    alpha, printable, avg, sent, upper, nav, repeat, nw = s.T\\n+    prior = np.zeros(len(ids))\\n+    prior += np.where((alpha > .62) & (alpha < .91), .25, -1.0)\\n+    prior += np.where((avg > 3.7) & (avg < 6.8), .15, -.5)\\n+    prior += np.where((sent > .025) & (sent < .16), .22, -.5)\\n+    prior += np.where(upper < .09, .12, -.35)\\n+    prior -= np.minimum(nav, .8) * .8 + np.maximum(repeat - .18, 0) * 2\\n+    prior += np.where((nw > 180) & (nw < 10_000), .12, -.2)\\n+    charlen, tech, html, news, wiki, bad, codechars, weird, moj = ex.T\\n+    quality = qscore + prior - .10 * np.log1p(bad) - .035 * moj - .035 * np.maximum(codechars - 5, 0)\\n+    quality -= .10 * np.maximum(np.log(np.maximum(lengths[ids], 1) / 8000), 0)\\n+    valid = ((lengths[ids] >= 100) & (lengths[ids] <= 10_000) & (alpha > .52) &\\n+             (printable > .985) & (avg > 3.2) & (avg < 7.5) & (upper < .20))\\n+    quality -= np.where(valid, 0, 5.0)\\n+\\n+    dom = probs.argmax(1)\\n+    lists = []\\n+    for k in range(4):\\n+        dk = np.where(dom == k)[0]\\n+        tb = .05 * np.log1p(tech[dk]) + .025 * np.log1p(html[dk])\\n+        bonus = [.06 * np.log1p(wiki[dk]), np.zeros(len(dk)), .04 * np.log1p(news[dk]), tb][k]\\n+        value = quality[dk] + .10 * np.log(probs[dk, k] + .03) + bonus\\n+        lists.append(dk[np.argsort(-value)].tolist())\\n+\\n+    weights = [1.60, 1.00, 1.00, 1.40]\\n+    ptr, load = [0] * 4, [0.0] * 4\\n+    used_ids, used_hashes, order = set(), set(), []\\n+    while len(order) < len(ids):\\n+        k = min(range(4), key=lambda j: load[j] / weights[j])\\n+        while ptr[k] < len(lists[k]):\\n+            ix = lists[k][ptr[k]]\\n+            if int(ids[ix]) not in used_ids and int(hashes[ix]) not in used_hashes:\\n+                break\\n+            ptr[k] += 1\\n+        if ptr[k] >= len(lists[k]):\\n+            load[k] = 1e30\\n+            if min(load) > 1e20:\\n+                break\\n+            continue\\n+        ix = lists[k][ptr[k]]\\n+        ptr[k] += 1\\n+        did, hh = int(ids[ix]), int(hashes[ix])\\n+        used_ids.add(did); used_hashes.add(hh); order.append(did)\\n+        load[k] += int(lengths[did])\\n+\\n+    Path(args.output).parent.mkdir(parents=True, exist_ok=True)\\n+    with open(args.output, \\\"w\\\") as f:\\n+        json.dump(order, f)\\n+    print(f\\\"wrote {len(order)} unique ids to {args.output}; total available tokens={sum(int(lengths[i]) for i in order)}\\\")\\n+\\n+\\n+if __name__ == \\\"__main__\\\":\\n+    main()\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:39081,chars:\"\",yield_time_ms:30000,max_output_tokens:12000});\ntext(r.output);\n"
}

{
  "cell_id": "155",
  "yield_time_ms": 30000,
  "max_tokens": 12000
}

{
  "input": "const r = await tools.write_stdin({session_id:39081,chars:\"\",yield_time_ms:30000,max_output_tokens:12000});\ntext(r.output);\n"
}

{
  "cell_id": "156",
  "yield_time_ms": 30000,
  "max_tokens": 12000
}

{
  "input": "const r = await tools.write_stdin({session_id:39081,chars:\"\",yield_time_ms:30000,max_output_tokens:12000});\ntext(r.output);\n"
}

{
  "cell_id": "157",
  "yield_time_ms": 30000,
  "max_tokens": 12000
}

{
  "input": "const r = await tools.write_stdin({session_id:39081,chars:\"\",yield_time_ms:30000,max_output_tokens:12000});\ntext(r.output);\n"
}

{
  "cell_id": "158",
  "yield_time_ms": 30000,
  "max_tokens": 12000
}

{
  "input": "const r = await tools.write_stdin({session_id:39081,chars:\"\",yield_time_ms:30000,max_output_tokens:12000});\ntext(r.output);\n"
}

{
  "cell_id": "159",
  "yield_time_ms": 30000,
  "max_tokens": 12000
}