gpt-5.6-sol

{
  "input": "const names=[\"minhash_news275_old\",\"minhash_info_mix_old\"];const rs=await Promise.all(names.map(n=>tools.exec_command({cmd:`python pack_selection.py candidates/${n}.json /tmp/${n}.npy`,workdir:\"/workspace\",yield_time_ms:1000,max_output_tokens:1200})));rs.forEach((r,i)=>text(names[i]+\": \"+JSON.stringify(r)));\n"
}

{
  "input": "const ids=[69145,65359];const rs=await Promise.all(ids.map(session_id=>tools.write_stdin({session_id,chars:\"\",yield_time_ms:30000,max_output_tokens:1200})));rs.forEach(r=>text(r));\n"
}

{
  "cell_id": "95",
  "yield_time_ms": 30000,
  "max_tokens": 4000
}

{
  "input": "const names=[\"minhash_news275_old\",\"minhash_info_mix_old\"];const cmd=names.map(n=>`python train_nano.py --train_npy /tmp/${n}.npy --heldout_npy data/multi_dev.npy --out_json /tmp/${n}_eval.json`).join(\" && \");const r=await tools.exec_command({cmd,workdir:\"/workspace\",yield_time_ms:1000,max_output_tokens:4000});text(r);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:19270,chars:\"\",yield_time_ms:30000,max_output_tokens:4000});\ntext(r);\n"
}

{
  "cell_id": "97",
  "yield_time_ms": 30000,
  "max_tokens": 4000
}

{
  "input": "const r = await tools.write_stdin({session_id:19270,chars:\"\",yield_time_ms:30000,max_output_tokens:4000});\ntext(r);\n"
}

{
  "cell_id": "98",
  "yield_time_ms": 30000,
  "max_tokens": 4000
}

{
  "input": "const r = await tools.write_stdin({session_id:19270,chars:\"\",yield_time_ms:30000,max_output_tokens:4000});\ntext(r);\n"
}

{
  "cell_id": "99",
  "yield_time_ms": 30000,
  "max_tokens": 4000
}

{
  "input": "const r = await tools.write_stdin({session_id:19270,chars:\"\",yield_time_ms:30000,max_output_tokens:4000});\ntext(r);\n"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Add File: /workspace/submission/curate.py\\n+#!/usr/bin/env python3\\n+\\\"\\\"\\\"Curate four balanced high-quality lanes with conservative near-deduplication.\\n+\\n+The disclosed dev corpus is used only as a style/domain reference: aggregate word\\n+statistics are computed separately for its four 250k-token quarters.  Pool pages\\n+are ranked by English prose quality plus affinity to encyclopedia, quality-web,\\n+news, or technical-Q&A language.  Priority prefixes contain equal estimated text\\n+mass from the four exclusive lanes.  Literal/whitespace duplicates and pages\\n+sharing at least three substantive sentence fingerprints are suppressed.\\n+\\\"\\\"\\\"\\n+import collections\\n+import hashlib\\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(\\\"/workspace\\\")\\n+POOL = ROOT / \\\"data/pool.jsonl\\\"\\n+DEV = ROOT / \\\"data/multi_dev.npy\\\"\\n+OUT = ROOT / \\\"submission/selection.json\\\"\\n+ALPHA_RE = re.compile(r\\\"[A-Za-z]+(?:['’][A-Za-z]+)?\\\")\\n+\\n+COMMON = set(\\\"\\\"\\\"the of and to in a is that for it as was with be by on are this from or\\n+at an not have has but they which you were their one all we can more there been if will\\n+would about who so when what its into no also than other some could them these may such\\n+only new time do two first any my now like our over after most me even because how many\\n+well where people through between had he she his her your i us out up get use\\\"\\\"\\\".split())\\n+TECH = set(\\\"\\\"\\\"code function class method variable object array string integer python java\\n+javascript linux windows database server client api application software program programming\\n+command file directory package module library framework error exception install configuration\\n+configure query sql html css algorithm data compiler memory network request response protocol\\n+interface type return value parameter output input github version system process thread shell\\n+terminal user implementation solution example using use problem issue test build runtime\\n+developer development web xml json\\\"\\\"\\\".split())\\n+NEWS = set(\\\"\\\"\\\"said says according reported reports officials government president minister\\n+police court state country monday tuesday wednesday thursday friday saturday sunday year percent\\n+million billion company group spokesman spokeswoman agency election law killed died attack\\n+military city international news market shares county council\\\"\\\"\\\".split())\\n+ENC = set(\\\"\\\"\\\"known called located founded species century history population region area\\n+including refers defined consists became developed established based form type example however\\n+during later early modern theory research study university language system include generally\\n+typically commonly number several may can also such first world\\\"\\\"\\\".split())\\n+BAD = (\\\"cookie policy\\\", \\\"all rights reserved\\\", \\\"privacy policy\\\", \\\"terms of use\\\", \\\"click here\\\",\\n+       \\\"sign up\\\", \\\"log in\\\", \\\"subscribe to\\\", \\\"javascript is required\\\", \\\"enable javascript\\\",\\n+       \\\"follow us on\\\", \\\"share this\\\", \\\"free shipping\\\", \\\"add to cart\\\", \\\"customer reviews\\\",\\n+       \\\"contact us\\\", \\\"read more\\\", \\\"view all\\\")\\n+\\n+# Aggregate reference distributions.  Keeping quarters separate prevents the easiest\\n+# or most common register from taking the whole 12M-token budget.\\n+tokenizer = AutoTokenizer.from_pretrained(\\\"gpt2\\\")\\n+dev_ids = np.load(DEV)\\n+domains = []\\n+for j in range(4):\\n+    text = tokenizer.decode(dev_ids[j * 250_000:(j + 1) * 250_000])\\n+    domains.append(collections.Counter(w.lower().replace(\\\"’\\\", \\\"'\\\") for w in ALPHA_RE.findall(text)))\\n+target = sum(domains, collections.Counter())\\n+target_total = sum(target.values())\\n+domain_totals = [sum(x.values()) for x in domains]\\n+target_vocab = {w for w, n in target.items() if n >= 2}\\n+word_stats = {}\\n+for w, total_n in target.items():\\n+    word_stats[w] = (\\n+        max(-12.0, math.log((total_n + .2) / target_total)),\\n+        tuple(math.log((dc.get(w, 0) + .2) / (total_n + .8)) +\\n+              math.log(target_total / dt) for dc, dt in zip(domains, domain_totals)),\\n+    )\\n+\\n+def score_document(doc):\\n+    text = doc[\\\"text\\\"]\\n+    low = text.lower()\\n+    words = [w.lower().replace(\\\"’\\\", \\\"'\\\") for w in ALPHA_RE.findall(text)]\\n+    n = len(words)\\n+    if not n:\\n+        return None\\n+    count = collections.Counter(words)\\n+    chars = len(text)\\n+    lines = [x.strip() for x in text.splitlines() if x.strip()] or [text]\\n+    short_line = sum(len(x) < 45 for x in lines) / len(lines)\\n+    line_unique = len(set(lines)) / len(lines)\\n+    common = sum(count[w] for w in COMMON) / n\\n+    vocab = sum(v for w, v in count.items() if w in target_vocab) / n\\n+    logp_sum = 0.0\\n+    affin_sum = [0.0] * 4\\n+    for w, v in count.items():\\n+        stats = word_stats.get(w)\\n+        if stats is None:\\n+            logp_sum -= 12.0 * v\\n+        else:\\n+            logp_sum += stats[0] * v\\n+            for j in range(4):\\n+                affin_sum[j] += stats[1][j] * v\\n+    logp = logp_sum / n\\n+    unique = len(count) / n\\n+    alpha_chars = sum(map(len, words))\\n+    upper = len(re.findall(r\\\"[A-Z]\\\", text)) / max(1, alpha_chars)\\n+    digits = len(re.findall(r\\\"[0-9]\\\", text)) / max(1, chars)\\n+    punct = len(re.findall(r\\\"[.,;:!?]\\\", text)) / max(1, chars)\\n+    bad = sum(low.count(x) for x in BAD)\\n+    urls = len(re.findall(r\\\"https?://|www\\\\.\\\", low))\\n+    weird = text.count(\\\"�\\\") + text.count(\\\"Â\\\") + text.count(\\\"Ã\\\")\\n+\\n+    quality = 3.2 * vocab + .42 * logp\\n+    quality -= 5.0 * abs(common - .37)\\n+    quality += .32 * min(math.log1p(n), 8.0)\\n+    quality += 1.4 * min(punct, .035) / .035\\n+    quality -= 2.0 * max(0.0, upper - .12) + 2.0 * max(0.0, digits - .12)\\n+    quality -= 1.3 * max(0.0, short_line - .55)\\n+    quality -= max(0.0, .7 - line_unique)\\n+    quality -= .08 * min(bad, 15) + .04 * min(urls, 20) + .25 * min(weird, 5)\\n+    quality -= 2.0 * max(0.0, .18 - unique)\\n+    if chars < 500:\\n+        quality -= (500 - chars) / 250\\n+    if chars > 50_000:\\n+        quality -= math.log(chars / 50_000)\\n+    if alpha_chars / max(1, chars) < .55:\\n+        quality -= 1.5\\n+    return {\\n+        \\\"id\\\": doc[\\\"id\\\"], \\\"chars\\\": chars, \\\"q\\\": quality,\\n+        \\\"tech\\\": sum(count[w] for w in TECH) / n,\\n+        \\\"news\\\": sum(count[w] for w in NEWS) / n,\\n+        \\\"enc\\\": sum(count[w] for w in ENC) / n,\\n+        \\\"aff\\\": [x / n for x in affin_sum],\\n+    }\\n+\\n+rows, fingerprints, format_features = [], {}, {}\\n+with open(POOL) as source:\\n+    for k, line in enumerate(source, 1):\\n+        doc = json.loads(line)\\n+        row = score_document(doc)\\n+        if row is None:\\n+            continue\\n+        rows.append(row)\\n+        text = doc[\\\"text\\\"]\\n+        longest = max(text.split(\\\"<|endoftext|>\\\"), key=len).strip()\\n+        exact = hashlib.blake2b(text.encode(), digest_size=12).hexdigest()\\n+        normalized = hashlib.blake2b(re.sub(r\\\"\\\\s+\\\", \\\" \\\", text).strip().encode(),\\n+                                     digest_size=12).hexdigest()\\n+        units = set()\\n+        for unit in re.split(r\\\"(?<=[.!?])\\\\s+|\\\\n+\\\", longest):\\n+            unit = re.sub(r\\\"\\\\s+\\\", \\\" \\\", unit.strip().lower())\\n+            if len(unit) >= 70:\\n+                units.add(hashlib.blake2b(unit.encode(), digest_size=8).hexdigest())\\n+        fingerprints[doc[\\\"id\\\"]] = (exact, normalized, tuple(sorted(units)[:6]))\\n+        format_features[doc[\\\"id\\\"]] = (\\n+            sum(x in text for x in (\\\"|Country||\\\", \\\"|Location||\\\", \\\"|Born||\\\", \\\"|Type||\\\",\\n+                                    \\\"|Kingdom||\\\", \\\"|Established||\\\")),\\n+            int(\\\"<pre><code>\\\" in text) + int(\\\"<code>\\\" in text) +\\n+                int(\\\"Posted on \\\" in text and \\\"?\\\" in text),\\n+            int(\\\"Reuters\\\" in text) + int(\\\"Associated Press\\\" in text) +\\n+                int(\\\" told \\\" in text.lower() and \\\" said \\\" in text.lower()),\\n+        )\\n+        if k % 25_000 == 0:\\n+            print(f\\\"scored {k} documents\\\", flush=True)\\n+\\n+def lane_scores(row):\\n+    wiki_fmt, qa_fmt, news_fmt = format_features[row[\\\"id\\\"]]\\n+    return (\\n+        row[\\\"q\\\"] + 10 * row[\\\"enc\\\"] + 1.5 * row[\\\"aff\\\"][0] + .15 * wiki_fmt,\\n+        row[\\\"q\\\"] + .6 * row[\\\"aff\\\"][1],\\n+        row[\\\"q\\\"] + 12 * row[\\\"news\\\"] + 1.2 * row[\\\"aff\\\"][2] + .08 * news_fmt,\\n+        row[\\\"q\\\"] + 15 * row[\\\"tech\\\"] + 1.5 * row[\\\"aff\\\"][3] + .12 * qa_fmt,\\n+    )\\n+\\n+# Assign each page to its most distinctive register, then rank within that lane.\\n+lanes = [[] for _ in range(4)]\\n+for row in rows:\\n+    scores = lane_scores(row)\\n+    lane = max(range(4), key=lambda j: scores[j] - row[\\\"q\\\"])\\n+    lanes[lane].append((scores[lane], row))\\n+for lane in lanes:\\n+    lane.sort(key=lambda item: item[0], reverse=True)\\n+\\n+# Weighted round robin by characters is a tokenizer-independent proxy for equal token\\n+# mass.  250 MB makes the ordered fallback much deeper than the 12M-token cutoff.\\n+positions = [0] * 4\\n+mass = [0] * 4\\n+selected, used, rejected = [], set(), []\\n+seen_exact, seen_normalized, seen_unit = set(), set(), {}\\n+while sum(mass) < 250_000_000:\\n+    active = [j for j in range(4) if positions[j] < len(lanes[j])]\\n+    if not active:\\n+        break\\n+    j = min(active, key=lambda x: mass[x])\\n+    row = lanes[j][positions[j]][1]\\n+    positions[j] += 1\\n+    doc_id = row[\\\"id\\\"]\\n+    if doc_id in used:\\n+        continue\\n+    exact, normalized, units = fingerprints[doc_id]\\n+    duplicate = exact in seen_exact or normalized in seen_normalized\\n+    if not duplicate and len(units) >= 3:\\n+        owners = collections.Counter(owner for unit in units for owner in seen_unit.get(unit, ()))\\n+        duplicate = bool(owners and max(owners.values()) >= 3)\\n+    if duplicate:\\n+        used.add(doc_id)\\n+        rejected.append(doc_id)\\n+        continue\\n+    used.add(doc_id)\\n+    selected.append(doc_id)\\n+    mass[j] += row[\\\"chars\\\"]\\n+    seen_exact.add(exact)\\n+    seen_normalized.add(normalized)\\n+    owner = len(selected)\\n+    for unit in units:\\n+        seen_unit.setdefault(unit, []).append(owner)\\n+\\n+# Valid deep fallback: quality-ranked unseen pages, then suppressed duplicates last.\\n+selected.extend(r[\\\"id\\\"] for r in sorted(rows, key=lambda x: x[\\\"q\\\"], reverse=True)\\n+                if r[\\\"id\\\"] not in used)\\n+selected.extend(rejected)\\n+OUT.parent.mkdir(parents=True, exist_ok=True)\\n+with open(OUT, \\\"w\\\") as f:\\n+    json.dump(selected, f)\\n+print(f\\\"wrote {len(selected)} unique ids to {OUT}; lane MB: \\\" +\\n+      \\\", \\\".join(f\\\"{x/1e6:.1f}\\\" for x in mass))\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\n"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Add File: /workspace/claim.md\\n+# Hypothesis\\n+\\n+At a 12M-token budget, a small model benefits more from diverse, target-shaped English than from either random web text or a single global “quality” ranking. The best subset should combine natural-prose filtering, explicit coverage of all four disclosed registers, and conservative removal of substantive near-duplicates.\\n+\\n+# Mechanism\\n+\\n+`curate.py` scores prose fluency and aggregate lexical affinity to each 250k-token dev quarter, assigns every document to its most distinctive register, and interleaves the four ranked lanes by equal character mass. It suppresses a candidate only when it is text-identical or shares at least three long sentence fingerprints with an earlier page.\\n+\\n+Observable predictions independent of final perplexity are: the consumed prefix will have approximately equal character mass from the four lanes; it will contain fewer repeated substantive sentences than an unfiltered prefix; and a model trained on it will improve loss in every dev quarter relative to random selection, rather than obtaining its gain from only one domain.\\n+\\n+# Falsification\\n+\\n+The hypothesis is falsified if a random or global-quality-only subset matches the four-lane subset, if deduplication does not reduce repeated sentence units, or if dev improvement is confined to one quarter while another materially regresses. It is also weakened if changing the dev sample changes selected registers drastically, which would indicate topic overfitting rather than style/quality selection.\\n+\\n+# Transfer\\n+\\n+The method transfers to another corpus by replacing the reference quarters and domain lexicons while retaining the quality features, exclusive lane assignment, mass-balanced scheduling, and sentence-unit deduplication. If no target sample is available, small representative reference corpora for the intended registers can substitute for the dev quarters.\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python submission/curate.py\",\"workdir\":\"/workspace\",\"yield_time_ms\":1000,\"max_output_tokens\":5000});\ntext(r);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:5978,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r);\n"
}

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

{
  "input": "const r = await tools.write_stdin({session_id:5978,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r);\n"
}

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

{
  "input": "const r = await tools.write_stdin({session_id:5978,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r);\n"
}

{
  "cell_id": "106",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:5978,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r);\n"
}

{
  "cell_id": "107",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:5978,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r);\n"
}

{
  "cell_id": "108",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}