gpt-5.6-sol

{
  "input": "const r = await tools.write_stdin({session_id:68624,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r.output);"
}

{
  "input": "const r = await tools.exec_command({cmd:\"find /root/.cache/huggingface/hub/models--gpt2 -maxdepth 3 -type f -printf '%p %s\\\\n'\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":5000});\ntext(r.output);"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Add File: /workspace/experiments/score_corpus.py\\n+import json, math, random, re, time\\n+from collections import Counter\\n+import numpy as np\\n+from transformers import GPT2TokenizerFast\\n+from sklearn.feature_extraction.text import TfidfVectorizer\\n+from sklearn.linear_model import LogisticRegression\\n+\\n+POOL = \\\"/workspace/data/pool.jsonl\\\"\\n+DEV = \\\"/workspace/data/multi_dev.npy\\\"\\n+OUT = \\\"/workspace/experiments/corpus_scores.npz\\\"\\n+random.seed(20260803)\\n+\\n+tok = GPT2TokenizerFast.from_pretrained(\\\"gpt2\\\", local_files_only=True)\\n+dev = np.load(DEV)\\n+\\n+def chunks(s, target=2500):\\n+    # Preserve whole target documents where practical, but split very long ones.\\n+    out = []\\n+    for d in s.split(\\\"<|endoftext|>\\\"):\\n+        d = d.strip()\\n+        if len(d) < 350:\\n+            continue\\n+        for st in range(0, len(d), target):\\n+            z = d[st:st+target]\\n+            if len(z) >= 350:\\n+                out.append(z)\\n+    return out\\n+\\n+train_x, train_y = [], []\\n+for k in range(4):\\n+    s = tok.decode(dev[k*250000:(k+1)*250000])\\n+    z = chunks(s)\\n+    train_x += z\\n+    train_y += [k] * len(z)\\n+    print(\\\"target\\\", k, len(z), flush=True)\\n+\\n+# Sample pool documents as the contrast class. The classifier then estimates a\\n+# density ratio between the disclosed high-quality target and raw pool.\\n+neg_ids = set(random.sample(range(182016), 12000))\\n+for line in open(POOL):\\n+    d = json.loads(line)\\n+    if d[\\\"id\\\"] in neg_ids:\\n+        z = d[\\\"text\\\"][:5000]\\n+        if len(z) >= 100:\\n+            train_x.append(z); train_y.append(4)\\n+\\n+vec = TfidfVectorizer(lowercase=True, strip_accents=\\\"unicode\\\", ngram_range=(1,2),\\n+                      min_df=2, max_df=.995, max_features=240000,\\n+                      sublinear_tf=True, norm=\\\"l2\\\", token_pattern=r\\\"(?u)\\\\b[\\\\w+#.-]{2,}\\\\b\\\")\\n+t0=time.time(); X=vec.fit_transform(train_x); print(\\\"vectorized\\\",X.shape,time.time()-t0,flush=True)\\n+clf = LogisticRegression(C=3.0, max_iter=150, solver=\\\"liblinear\\\", multi_class=\\\"ovr\\\",\\n+                         class_weight=\\\"balanced\\\", n_jobs=1)\\n+clf.fit(X, train_y); print(\\\"fit\\\",time.time()-t0,flush=True)\\n+\\n+names = [\\\"id\\\",\\\"nchar\\\",\\\"nword\\\",\\\"alpha\\\",\\\"digit\\\",\\\"newline\\\",\\\"eot\\\",\\\"url\\\",\\\"html\\\",\\n+         \\\"nav\\\",\\\"repeat\\\",\\\"quality\\\",\\\"d0\\\",\\\"d1\\\",\\\"d2\\\",\\\"d3\\\",\\\"junk\\\"]\\n+cols = {n: [] for n in names}\\n+\\n+nav_re = re.compile(r\\\"\\\\b(?:login|register|menu|cookie|privacy policy|terms and conditions|\\\"\\n+                    r\\\"skip to content|my account|shopping cart|subscribe|newsletter|\\\"\\n+                    r\\\"javascript|all rights reserved|contact us|home page|search results)\\\\b\\\", re.I)\\n+url_re = re.compile(r\\\"https?://|www\\\\.|\\\\.com\\\\b\\\", re.I)\\n+html_re = re.compile(r\\\"</?(?:p|div|span|a|li|pre|code|table|script)\\\\b|&(?:gt|lt|nbsp|quot);\\\",re.I)\\n+\\n+def meta(d):\\n+    s=d[\\\"text\\\"]; n=max(1,len(s)); words=re.findall(r\\\"[A-Za-z]+(?:'[A-Za-z]+)?\\\",s)\\n+    nw=max(1,len(words)); lower=[w.lower() for w in words]\\n+    grams=list(zip(lower,lower[1:],lower[2:]))\\n+    rep=1-len(set(grams))/max(1,len(grams))\\n+    nav=len(nav_re.findall(s))/nw\\n+    alpha=sum(c.isalpha() for c in s)/n\\n+    digit=sum(c.isdigit() for c in s)/n\\n+    nl=s.count(\\\"\\\\n\\\")/n\\n+    urls=len(url_re.findall(s))/nw\\n+    html=len(html_re.findall(s))/nw\\n+    # Transparent baseline quality feature, separate from density-ratio score.\\n+    q=(2.5*min(alpha,.82)-12*abs(digit-.025)-20*nav-8*urls-5*max(0,rep-.08)\\n+       -12*max(0,nl-.035)-0.25*abs(math.log(max(250,n)/3000)))\\n+    return n,nw,alpha,digit,nl,s.count(\\\"<|endoftext|>\\\"),urls,html,nav,rep,q\\n+\\n+buf=[]; bmeta=[]\\n+def flush():\\n+    if not buf:return\\n+    P=clf.predict_proba(vec.transform([s[:5000] for s in buf]))\\n+    for d,m,p in zip(bmeta,[meta(d) for d in bmeta],P):\\n+        vals=[d[\\\"id\\\"],*m,*p.tolist()]\\n+        for name,val in zip(names,vals): cols[name].append(val)\\n+    buf.clear();bmeta.clear()\\n+\\n+t1=time.time()\\n+for line in open(POOL):\\n+    d=json.loads(line);buf.append(d[\\\"text\\\"]);bmeta.append(d)\\n+    if len(buf)>=1024: flush()\\n+    if d[\\\"id\\\"] and d[\\\"id\\\"]%20000==0: print(\\\"score\\\",d[\\\"id\\\"],round(time.time()-t1,1),flush=True)\\n+flush()\\n+np.savez(OUT,**{k:np.asarray(v,dtype=np.int32 if k in (\\\"id\\\",\\\"nchar\\\",\\\"nword\\\",\\\"eot\\\") else np.float32) for k,v in cols.items()})\\n+print(\\\"saved\\\",OUT,len(cols[\\\"id\\\"]),round(time.time()-t1,1),flush=True)\\n*** End Patch\";\ntext(await tools.apply_patch(patch));"
}

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

{
  "input": "const r = await tools.write_stdin({session_id:83822,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r.output);"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python - <<'PY'\\nfor x in ['scipy','pandas','xgboost','lightgbm','spacy']:\\n try:\\n  m=__import__(x); print(x,'yes',getattr(m,'__version__',''))\\n except Exception as e: print(x,'no')\\nPY\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":3000});\ntext(r.output);"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Update File: /workspace/experiments/score_corpus.py\\n@@\\n-from sklearn.feature_extraction.text import TfidfVectorizer\\n-from sklearn.linear_model import LogisticRegression\\n@@\\n-train_x, train_y = [], []\\n+train_x, train_y = [], []\\n@@\\n-vec = TfidfVectorizer(lowercase=True, strip_accents=\\\"unicode\\\", ngram_range=(1,2),\\n-                      min_df=2, max_df=.995, max_features=240000,\\n-                      sublinear_tf=True, norm=\\\"l2\\\", token_pattern=r\\\"(?u)\\\\b[\\\\w+#.-]{2,}\\\\b\\\")\\n-t0=time.time(); X=vec.fit_transform(train_x); print(\\\"vectorized\\\",X.shape,time.time()-t0,flush=True)\\n-clf = LogisticRegression(C=3.0, max_iter=150, solver=\\\"liblinear\\\", multi_class=\\\"ovr\\\",\\n-                         class_weight=\\\"balanced\\\", n_jobs=1)\\n-clf.fit(X, train_y); print(\\\"fit\\\",time.time()-t0,flush=True)\\n+word_re = re.compile(r\\\"[a-zA-Z][a-zA-Z0-9_+#.-]*|[0-9]+|</?\\\\w+>|&\\\\w+;\\\")\\n+counts = [[Counter(), Counter()] for _ in range(5)]\\n+totals = [[0, 0] for _ in range(5)]\\n+for s,y in zip(train_x,train_y):\\n+    w=[x.lower() for x in word_re.findall(s)]\\n+    g=[w[i]+\\\"\\\\x01\\\"+w[i+1] for i in range(len(w)-1)]\\n+    counts[y][0].update(w); counts[y][1].update(g)\\n+for y in range(5):\\n+    for o in range(2): totals[y][o]=sum(counts[y][o].values())\\n+print(\\\"counted\\\",[(totals[y],len(counts[y][0]),len(counts[y][1])) for y in range(5)],flush=True)\\n+\\n+# Additive smoothing is computed over the union vocabulary.  Lookup tables hold\\n+# each feature's log probability under every class.\\n+logs=[]; defaults=[]\\n+for o in range(2):\\n+    vocab=set().union(*(counts[y][o] for y in range(5))); V=len(vocab); alpha=.1\\n+    den=[math.log(totals[y][o]+alpha*V) for y in range(5)]\\n+    defaults.append([math.log(alpha)-den[y] for y in range(5)])\\n+    logs.append({x:tuple(math.log(counts[y][o].get(x,0)+alpha)-den[y] for y in range(5)) for x in vocab})\\n+    print(\\\"vocab\\\",o,V,flush=True)\\n+\\n+def density(s):\\n+    w=[x.lower() for x in word_re.findall(s[:5000])]\\n+    feats=(w,[w[i]+\\\"\\\\x01\\\"+w[i+1] for i in range(len(w)-1)])\\n+    ans=np.zeros(5,dtype=np.float64)\\n+    for o,z in enumerate(feats):\\n+        if not z: continue\\n+        sm=np.zeros(5)\\n+        for x in z: sm += logs[o].get(x,defaults[o])\\n+        ans += (.35 if o==0 else .65)*sm/len(z)\\n+    # Convert cross-entropies to a stable relative score. Values are log-density\\n+    # advantages over the random raw-web contrast class, by target domain.\\n+    return ans[:4]-ans[4]\\n@@\\n-         \\\"nav\\\",\\\"repeat\\\",\\\"quality\\\",\\\"d0\\\",\\\"d1\\\",\\\"d2\\\",\\\"d3\\\",\\\"junk\\\"]\\n+         \\\"nav\\\",\\\"repeat\\\",\\\"quality\\\",\\\"d0\\\",\\\"d1\\\",\\\"d2\\\",\\\"d3\\\"]\\n@@\\n-    P=clf.predict_proba(vec.transform([s[:5000] for s in buf]))\\n-    for d,m,p in zip(bmeta,[meta(d) for d in bmeta],P):\\n+    P=[density(s) for s in buf]\\n+    for d,m,p in zip(bmeta,[meta(d) for d in bmeta],P):\\n*** End Patch\";\ntext(await tools.apply_patch(patch));"
}

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

{
  "input": "const r = await tools.write_stdin({session_id:66957,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r.output);"
}

{
  "cell_id": "25",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:66957,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r.output);"
}

{
  "cell_id": "26",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:66957,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r.output);"
}

{
  "cell_id": "27",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:66957,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r.output);"
}

{
  "cell_id": "28",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:66957,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r.output);"
}

{
  "cell_id": "29",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:66957,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r.output);"
}

{
  "cell_id": "30",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:66957,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r.output);"
}

{
  "cell_id": "31",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({session_id:66957,chars:\"\",yield_time_ms:30000,max_output_tokens:5000});\ntext(r.output);"
}