gpt-5.6-sol

{
  "input": "const r = await tools.write_stdin({session_id:42083,chars:\"\",yield_time_ms:1000,max_output_tokens:1000});\ntext(JSON.stringify(r));"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python train_diag.py --train_npy /tmp/mixE.npy --heldout_npy data/multi_dev.npy\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":2000});\ntext(JSON.stringify(r));"
}

{
  "cell_id": "75",
  "yield_time_ms": 30000,
  "max_tokens": 2000
}

{
  "input": "const r = await tools.write_stdin({session_id:77683,chars:\"\",yield_time_ms:30000,max_output_tokens:2000});\ntext(JSON.stringify(r));"
}

{
  "cell_id": "76",
  "yield_time_ms": 30000,
  "max_tokens": 2000
}

{
  "input": "const r = await tools.exec_command({cmd:\"python - <<'PY'\\nimport json,numpy as np,random\\nz=np.load('/tmp/pool_scores.npz');s=z['scores'];q=z['quality'];c=z['chars'];nw=z['nwords']\\nsel=json.load(open('/tmp/mixB.json')); used=sel[:15384]\\nprint('used n',len(used),'chars',sum(c[used]),'regions',sum(i<100000 for i in used),'q pct',np.percentile(q[used],[0,10,25,50,75,90,100]))\\nprint('score means selected',s[used].mean(0),'random',s[np.random.default_rng(1).choice(182016,len(used),False)].mean(0))\\n# Print evenly spaced selected samples.\\nwant={used[x] for x in range(0,len(used),1200)}\\nfor line in open('data/pool.jsonl'):\\n o=json.loads(line);i=o['id']\\n if i in want: print('\\\\n',used.index(i),i,'q',round(float(q[i]),2),'s',np.round(s[i],2),repr(' '.join(o['text'].split())[:400]))\\nPY\\ncat .timer/remaining_secs\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":10000});\ntext(r.output);"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Add File: /workspace/submission/curate.py\\n+#!/usr/bin/env python3\\n+\\\"\\\"\\\"Target-density-ratio curation with register and quality control.\\n+\\n+The disclosed dev target supplies four *distribution prototypes*, not document\\n+IDs or labels for the pool.  For each prototype (Wikipedia, quality web, news,\\n+technical Q&A), this script estimates a smoothed unigram density ratio against\\n+the entire raw pool.  Documents are ranked by mean clipped log density ratio.\\n+\\n+A language/format quality gate rejects short, non-prose, boilerplate-heavy,\\n+non-English, and highly templated pages before ranking.  Finally, a weighted\\n+fair queue interleaves the four rankings by estimated token mass.  The shares\\n+25/20/25/30 were selected before finalization from small dev ablations; the\\n+technical share is slightly larger because it showed less cross-register\\n+transfer than the three prose registers at this model scale.\\n+\\\"\\\"\\\"\\n+import collections\\n+import json\\n+import math\\n+import re\\n+from pathlib import Path\\n+\\n+import numpy as np\\n+from transformers import AutoTokenizer\\n+\\n+ROOT = Path(__file__).resolve().parents[1]\\n+POOL = ROOT / \\\"data\\\" / \\\"pool.jsonl\\\"\\n+TARGET = ROOT / \\\"data\\\" / \\\"multi_dev.npy\\\"\\n+OUT = Path(__file__).resolve().parent / \\\"selection.json\\\"\\n+\\n+WORD_RE = re.compile(r\\\"[a-z]+(?:'[a-z]+)?\\\")\\n+STOP = set((\\n+    \\\"the of and to a in is it that for on was with as are be at by this from or an \\\"\\n+    \\\"have not but they you which his her their has had were will would can one all we \\\"\\n+    \\\"about there if more when who what so no been do into than its also may some other \\\"\\n+    \\\"how these our your my i he she them\\\"\\n+).split())\\n+BAD = (\\n+    \\\"privacy policy\\\", \\\"cookie policy\\\", \\\"all rights reserved\\\", \\\"skip to content\\\",\\n+    \\\"javascript is disabled\\\", \\\"enable javascript\\\", \\\"log in register\\\",\\n+    \\\"sign in register\\\", \\\"terms and conditions\\\", \\\"buy now\\\", \\\"free shipping\\\",\\n+    \\\"shopping cart\\\", \\\"click here\\\", \\\"subscribe now\\\", \\\"online casino\\\",\\n+    \\\"payday loan\\\", \\\"erectile dysfunction\\\", \\\"weight loss pills\\\", \\\"viagra\\\", \\\"porn\\\",\\n+    \\\"categories archives\\\", \\\"menu home\\\", \\\"search this site\\\", \\\"website uses cookies\\\",\\n+    \\\"page not found\\\", \\\"access denied\\\",\\n+)\\n+N_DOCS = 182_016\\n+\\n+\\n+def words(text):\\n+    return WORD_RE.findall(text.lower())\\n+\\n+\\n+def target_profiles():\\n+    tokenizer = AutoTokenizer.from_pretrained(\\\"gpt2\\\", local_files_only=True)\\n+    ids = np.load(TARGET)\\n+    profiles = []\\n+    for k in range(4):\\n+        text = tokenizer.decode(ids[k * 250_000:(k + 1) * 250_000]).lower()\\n+        profiles.append(collections.Counter(WORD_RE.findall(text)))\\n+    aggregate = sum(profiles, collections.Counter())\\n+    # Singletons are mostly names/topics from this particular dev sample.  Dropping\\n+    # them makes the criterion about register rather than memorizing those topics.\\n+    vocab = {w for w, count in aggregate.items() if count >= 2}\\n+    return profiles, vocab\\n+\\n+\\n+def pool_background(vocab):\\n+    counts = collections.Counter()\\n+    total = 0\\n+    with POOL.open() as source:\\n+        for line in source:\\n+            ws = words(json.loads(line)[\\\"text\\\"])\\n+            total += len(ws)\\n+            counts.update(w for w in ws if w in vocab)\\n+    return counts, total\\n+\\n+\\n+def density_weights(profiles, vocab, background, background_total):\\n+    # Empirical-Bayes shrinkage prevents rare dev words from dominating a rank.\\n+    alpha = 10_000.0\\n+    unknown_count = background_total - sum(background.values())\\n+    result = []\\n+    for profile in profiles:\\n+        target_total = sum(profile.values())\\n+        weight = {}\\n+        for w in vocab:\\n+            raw_p = max(background[w] / background_total, 1 / background_total)\\n+            target_p = (profile[w] + alpha * raw_p) / (target_total + alpha)\\n+            weight[w] = float(np.clip(math.log(target_p / raw_p), -3.0, 3.0))\\n+        raw_p = unknown_count / background_total\\n+        target_p = alpha * raw_p / (target_total + alpha)\\n+        weight[None] = float(np.clip(math.log(target_p / raw_p), -3.0, 3.0))\\n+        result.append(weight)\\n+    return result\\n+\\n+\\n+def score_pool(vocab, weights):\\n+    scores = np.empty((N_DOCS, 4), dtype=np.float32)\\n+    quality = np.empty(N_DOCS, dtype=np.float32)\\n+    chars = np.empty(N_DOCS, dtype=np.int32)\\n+    nwords = np.empty(N_DOCS, dtype=np.int32)\\n+\\n+    with POOL.open() as source:\\n+        for line in source:\\n+            row = json.loads(line)\\n+            i, text = row[\\\"id\\\"], row[\\\"text\\\"]\\n+            low = text.lower()\\n+            ws = WORD_RE.findall(low)\\n+            nw, length = len(ws), max(1, len(text))\\n+            chars[i], nwords[i] = length, nw\\n+            if nw:\\n+                for k, weight in enumerate(weights):\\n+                    scores[i, k] = sum(weight.get(w, weight[None]) for w in ws) / nw\\n+                known = sum(w in vocab for w in ws) / nw\\n+                stop = sum(w in STOP for w in ws) / nw\\n+                average_word_length = sum(map(len, ws)) / nw\\n+            else:\\n+                scores[i, :] = -3.0\\n+                known = stop = 0.0\\n+                average_word_length = 20.0\\n+\\n+            alpha_fraction = sum(c.isalpha() for c in text) / length\\n+            digit_fraction = sum(c.isdigit() for c in text) / length\\n+            lines = text.splitlines() or [\\\"\\\"]\\n+            short_line_fraction = sum(len(x.strip()) < 40 for x in lines) / len(lines)\\n+            sentence_rate = (text.count(\\\".\\\") + text.count(\\\"!\\\") + text.count(\\\"?\\\")) / max(1, nw) * 100\\n+            bad = sum(marker in low for marker in BAD)\\n+            q = (\\n+                2.0 * known\\n+                + 1.2 * min(stop / 0.32, 1.0)\\n+                + 0.8 * min(alpha_fraction / 0.72, 1.0)\\n+                + 0.25 * min(sentence_rate / 3.0, 1.0)\\n+                - 0.8 * max(0.0, short_line_fraction - 0.35)\\n+                - 0.18 * bad\\n+                - 0.8 * max(0.0, digit_fraction - 0.12)\\n+                - 0.12 * abs(average_word_length - 4.8)\\n+            )\\n+            if nw < 120:\\n+                q -= (120 - nw) / 120\\n+            if length > 30_000:\\n+                q -= min(1.0, (length - 30_000) / 50_000)\\n+            quality[i] = q\\n+    return scores, quality, chars, nwords\\n+\\n+\\n+def fair_rank(scores, quality, chars, nwords):\\n+    eligible = (\\n+        (nwords >= 150) & (chars >= 700) & (chars <= 30_000) & (quality >= 3.5)\\n+    )\\n+    ids = np.flatnonzero(eligible)\\n+    rankings = [ids[np.argsort(-scores[ids, k])].tolist() for k in range(4)]\\n+\\n+    shares = (0.25, 0.20, 0.25, 0.30)\\n+    positions = [0, 0, 0, 0]\\n+    mass = [0.0, 0.0, 0.0, 0.0]\\n+    used, output = set(), []\\n+    # Far more than the 12M-token prefix; the verifier may safely stop early.\\n+    while sum(mass) < 40_000_000:\\n+        domain = min(range(4), key=lambda k: mass[k] / shares[k])\\n+        ranking = rankings[domain]\\n+        while ranking[positions[domain]] in used:\\n+            positions[domain] += 1\\n+        doc_id = ranking[positions[domain]]\\n+        positions[domain] += 1\\n+        used.add(doc_id)\\n+        output.append(int(doc_id))\\n+        # Characters/4 is used only for fair interleaving, never validity accounting.\\n+        mass[domain] += chars[doc_id] / 4 + 1\\n+    return output\\n+\\n+\\n+def main():\\n+    profiles, vocab = target_profiles()\\n+    background, total = pool_background(vocab)\\n+    weights = density_weights(profiles, vocab, background, total)\\n+    scores, quality, chars, nwords = score_pool(vocab, weights)\\n+    selection = fair_rank(scores, quality, chars, nwords)\\n+    OUT.write_text(json.dumps(selection) + \\\"\\\\n\\\")\\n+    print(f\\\"wrote {len(selection):,} unique ranked ids to {OUT}\\\")\\n+\\n+\\n+if __name__ == \\\"__main__\\\":\\n+    main()\\n*** End Patch\";\ntext(await tools.apply_patch(patch));"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Add File: /workspace/claim.md\\n+# Hypothesis\\n+\\n+Under a severe token budget, training on English documents with high smoothed\\n+log-density ratio to each target register, subject to a prose-quality floor and\\n+token-mass balancing, will outperform random web sampling. A modest technical\\n+oversample should help because code/Q&A transfers less readily from ordinary\\n+prose than the three prose registers transfer among themselves.\\n+\\n+# Mechanism\\n+\\n+The selected 12M-token prefix should have (1) a higher mean held-out\\n+target-versus-pool unigram log-density ratio in every register, (2) fewer short,\\n+non-English, boilerplate-heavy and navigation-heavy documents, and (3) estimated\\n+register mass near 25% encyclopedic, 20% general web, 25% news and 30% technical.\\n+These are observable without training the final model. The mechanism is reduced\\n+gradient budget spent modeling raw-web artifacts, while retaining vocabulary and\\n+style coverage for all four evaluation registers.\\n+\\n+# Falsification\\n+\\n+The hypothesis is falsified if the packed prefix does not improve those three\\n+pretraining-data observables over an equal-size random prefix, or if a controlled\\n+ablation matched for length and quality but shuffled across density-ratio ranks\\n+does not worsen per-register development cross-entropy. It would also be\\n+falsified if removing the fair queue improves all four register losses, since\\n+that would show balancing is unnecessary rather than causal.\\n+\\n+# Transfer\\n+\\n+The method transfers to another raw corpus when a small representative target\\n+sample exists: estimate target/raw density ratios, apply language-appropriate\\n+quality signals, and fair-queue the required registers. The English stopword and\\n+word-tokenization quality gate and the 25/20/25/30 shares are not expected to\\n+transfer unchanged across languages, tokenizers, model sizes, or target mixes;\\n+those components should be recalibrated while the density-ratio principle stays\\n+the same.\\n*** End Patch\";\ntext(await tools.apply_patch(patch));"
}

{
  "input": "const r = await tools.update_plan({explanation:\"Corpus profiling and target-aware ablations are complete; the best validated mixture is selected. I am now reproducing it from the publishable script and validating exact token coverage and files.\",plan:[\n  {step:\"Inspect corpus, evaluator, token accounting, and time budget\",status:\"completed\"},\n  {step:\"Profile document sources/quality signals and build reproducible ranker\",status:\"completed\"},\n  {step:\"Generate and benchmark candidate selections\",status:\"completed\"},\n  {step:\"Finalize selection, curate.py, claim.md, and validate deliverables\",status:\"in_progress\"}\n]});\nconst e = await tools.exec_command({cmd:\"python submission/curate.py\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":3000});\ntext(JSON.stringify(e));"
}

{
  "cell_id": "80",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.write_stdin({session_id:36617,chars:\"\",yield_time_ms:30000,max_output_tokens:3000});\ntext(JSON.stringify(r));"
}

{
  "cell_id": "81",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.write_stdin({session_id:36617,chars:\"\",yield_time_ms:30000,max_output_tokens:3000});\ntext(JSON.stringify(r));"
}

{
  "cell_id": "82",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.write_stdin({session_id:36617,chars:\"\",yield_time_ms:30000,max_output_tokens:3000});\ntext(JSON.stringify(r));"
}

{
  "cell_id": "83",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.write_stdin({session_id:36617,chars:\"\",yield_time_ms:30000,max_output_tokens:3000});\ntext(JSON.stringify(r));"
}

{
  "cell_id": "84",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.write_stdin({session_id:36617,chars:\"\",yield_time_ms:30000,max_output_tokens:3000});\ntext(JSON.stringify(r));"
}

{
  "cell_id": "85",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.exec_command({cmd:\"cmp -s submission/selection.json /tmp/mixB.json; echo cmp_exit=$?; python - <<'PY'\\nimport json,os\\np='submission/selection.json';x=json.load(open(p))\\nprint('count',len(x),'unique',len(set(x)),'min',min(x),'max',max(x),'bytes',os.path.getsize(p))\\ny=json.load(open('/tmp/mixB.json'))\\nprint('same list',x==y,'first10',x[:10])\\nPY\\npython fastpack.py submission/selection.json /tmp/final.npy\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":3000});\ntext(JSON.stringify(r));"
}

{
  "cell_id": "86",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.exec_command({cmd:\"python train_nano.py --train_npy /tmp/final.npy --heldout_npy data/multi_dev.npy --out_json /tmp/final_eval.json --max_iters 3000 --seed 1337\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":2000});\ntext(JSON.stringify(r));"
}