gpt-5.6-sol

{
  "input": "const r = await tools.exec_command({cmd:\"python build_candidate.py --weight 3 --shares .2167,.2167,.2166,.35 --name r3_tech35 && python build_candidate.py --weight 3 --shares .30,.20,.20,.30 --name r3_edges30 && python build_candidate.py --weight 3 --shares .20,.25,.25,.30 --name r3_tech30 && python build_candidate.py --weight 3 --shares .30,.2333,.2333,.2034 --name r3_wiki30\",workdir:\"/workspace\",yield_time_ms:30000,max_output_tokens:4000});\ntext(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python - <<'PY'\\nimport numpy as np,glob,os\\nD=np.load('data/multi_dev.npy');V=50257\\nnames=['ratio30','r3_tech35','r3_edges30','r3_tech30','r3_wiki30','r3b025','r3b05','r4b05']\\nfor name in names:\\n a=np.load('/tmp/'+name+'.npy').astype(int);c=np.bincount(a,minlength=V)+.1;p=c/c.sum();x=[]\\n for q in range(4):\\n  d=np.bincount(D[q*250000:(q+1)*250000],minlength=V)+.1;d=d/d.sum();x.append(np.sum(d*np.log(d/p)))\\n print(name, np.round(x,3),round(sum(x)/4,3))\\nPY\",workdir:\"/workspace\",yield_time_ms:10000,max_output_tokens:3000});\ntext(r.output);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:46449,chars:\"\",yield_time_ms:30000,max_output_tokens:2000});\ntext(JSON.stringify(r));\n"
}

{
  "cell_id": "97",
  "yield_time_ms": 30000,
  "max_tokens": 2000
}

{
  "input": "const r = await tools.write_stdin({session_id:46449,chars:\"\",yield_time_ms:30000,max_output_tokens:2000});\ntext(JSON.stringify(r));\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python train_nano.py --train_npy /tmp/ratio35.npy --heldout_npy data/multi_dev.npy --out_json /tmp/ratio35_result.json\",workdir:\"/workspace\",yield_time_ms:1000,max_output_tokens:1000});\ntext(JSON.stringify(r));\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python - <<'PY'\\nimport json,hashlib,collections\\nsels={x:json.load(open('/tmp/'+x+'.json'))[:14000] for x in ['ratio30','ratio35','r3b025']};sets={x:set(v) for x,v in sels.items()};H={x:[] for x in sels}; texts={}\\nfor l in open('data/pool.jsonl'):\\n d=json.loads(l);i=d['id']\\n for x,s in sets.items():\\n  if i in s:H[x].append(hashlib.blake2b(d['text'].strip().lower().encode(),digest_size=8).digest())\\nfor x,h in H.items():print(x,len(h),len(set(h)),'duplicates',len(h)-len(set(h)))\\nPY\",workdir:\"/workspace\",yield_time_ms:10000,max_output_tokens:1000});\ntext(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python - <<'PY'\\nimport json,hashlib,numpy as np\\nh=[]\\nfor l in open('data/pool.jsonl'):\\n t=json.loads(l)['text'].strip().lower().encode();h.append(int.from_bytes(hashlib.blake2b(t,digest_size=8).digest(),'little'))\\nnp.save('/tmp/text_hash.npy',np.asarray(h,dtype=np.uint64));print(len(h),len(set(h)))\\nPY\",workdir:\"/workspace\",yield_time_ms:10000,max_output_tokens:1000});\ntext(r.output);\n"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Update File: /workspace/build_candidate.py\\n@@\\n ap.add_argument(\\\"--shares\\\", default=\\\".25,.25,.25,.25\\\")\\n+ap.add_argument(\\\"--dedup\\\", action=\\\"store_true\\\")\\n@@\\n used = np.zeros(n, dtype=bool); selected=[]\\n+hashes = np.load('/tmp/text_hash.npy') if a.dedup else None\\n+used_hashes = set()\\n while used_tokens.sum() < 12_500_000:\\n@@\\n     i = int(orders[q][pos[q]]); pos[q] += 1\\n@@\\n     if quality[i] < -2.0: continue\\n+    if a.dedup and int(hashes[i]) in used_hashes: continue\\n     used[i] = True; selected.append(i); used_tokens[q] += lengths[i]\\n+    if a.dedup: used_hashes.add(int(hashes[i]))\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python build_candidate.py --weight 3 --dedup --name r3dedup\",workdir:\"/workspace\",yield_time_ms:30000,max_output_tokens:1000});\ntext(r.output);\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python - <<'PY'\\nimport numpy as np\\nD=np.load('data/multi_dev.npy')\\nfor q in range(4):\\n a=D[q*250000:(q+1)*250000];print(q,(a==50256).sum(),len(a)/max(1,(a==50256).sum()))\\nPY\",workdir:\"/workspace\",yield_time_ms:10000,max_output_tokens:1000});\ntext(r.output);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:99573,chars:\"\",yield_time_ms:30000,max_output_tokens:2000});\ntext(JSON.stringify(r));\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python train_nano.py --train_npy /tmp/r3dedup.npy --heldout_npy data/multi_dev.npy --out_json /tmp/r3dedup_result.json\",workdir:\"/workspace\",yield_time_ms:1000,max_output_tokens:1000});\ntext(JSON.stringify(r));\n"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Add File: /workspace/train_diag.py\\n+\\\"\\\"\\\"Diagnostic clone of frozen training recipe, with per-quarter dev losses.\\\"\\\"\\\"\\n+import argparse, math, time, json, numpy as np, torch\\n+from model import GPT, GPTConfig\\n+ap=argparse.ArgumentParser();ap.add_argument('--train_npy',required=True);ap.add_argument('--heldout_npy',required=True);ap.add_argument('--out_json',required=True);ap.add_argument('--max_iters',type=int,default=3000);ap.add_argument('--warmup',type=int,default=150);ap.add_argument('--batch',type=int,default=32);ap.add_argument('--block',type=int,default=256);ap.add_argument('--lr',type=float,default=6e-4);ap.add_argument('--seed',type=int,default=1337);a=ap.parse_args()\\n+torch.manual_seed(a.seed);np.random.seed(a.seed);dev='cuda'\\n+tr=torch.from_numpy(np.load(a.train_npy).astype(np.int64));ho=torch.from_numpy(np.load(a.heldout_npy).astype(np.int64));rng=np.random.default_rng(a.seed)\\n+model=GPT(GPTConfig(block_size=a.block,vocab_size=50257,n_layer=6,n_head=6,n_embd=384,dropout=0.0,bias=False)).to(dev);opt=model.configure_optimizers(.1,a.lr,(.9,.95),'cuda')\\n+def lr_at(it):\\n+ if it<a.warmup:return a.lr*(it+1)/(a.warmup+1)\\n+ r=(it-a.warmup)/max(1,a.max_iters-a.warmup);return .1*a.lr+.5*(1+math.cos(math.pi*r))*(a.lr-.1*a.lr)\\n+def batch(src):\\n+ ix=rng.integers(0,len(src)-a.block-1,size=a.batch);x=torch.stack([src[i:i+a.block] for i in ix]).to(dev);y=torch.stack([src[i+1:i+1+a.block] for i in ix]).to(dev);return x,y\\n+model.train()\\n+for it in range(a.max_iters):\\n+ for g in opt.param_groups:g['lr']=lr_at(it)\\n+ x,y=batch(tr)\\n+ with torch.autocast('cuda',dtype=torch.bfloat16):_,loss=model(x,y)\\n+ opt.zero_grad(set_to_none=True);loss.backward();torch.nn.utils.clip_grad_norm_(model.parameters(),1.0);opt.step()\\n+def evaluate(src):\\n+ nwin=(len(src)-1)//a.block;total=0\\n+ with torch.no_grad():\\n+  for j in range(0,nwin,a.batch):\\n+   ss=list(range(j,min(j+a.batch,nwin)));x=torch.stack([src[k*a.block:k*a.block+a.block] for k in ss]).to(dev);y=torch.stack([src[k*a.block+1:k*a.block+1+a.block] for k in ss]).to(dev)\\n+   with torch.autocast('cuda',dtype=torch.bfloat16):_,l=model(x,y)\\n+   total+=l.item()*len(ss)\\n+ return total/nwin\\n+model.eval();result={'all':evaluate(ho)}\\n+for q in range(4):result[str(q)]=evaluate(ho[q*250000:(q+1)*250000])\\n+print(json.dumps(result));json.dump(result,open(a.out_json,'w'))\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python -m py_compile train_diag.py\",workdir:\"/workspace\",yield_time_ms:10000,max_output_tokens:1000});\ntext(r.output);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:49653,chars:\"\",yield_time_ms:30000,max_output_tokens:2000});\ntext(JSON.stringify(r));\n"
}

{
  "cell_id": "109",
  "yield_time_ms": 30000,
  "max_tokens": 2000
}

{
  "input": "const r = await tools.exec_command({cmd:\"python train_diag.py --train_npy /tmp/ratio30.npy --heldout_npy data/multi_dev.npy --out_json /tmp/ratio30_diag.json\",workdir:\"/workspace\",yield_time_ms:1000,max_output_tokens:1000});\ntext(JSON.stringify(r));\n"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Update File: /workspace/build_candidate.py\\n@@\\n ap.add_argument(\\\"--weight\\\", type=float, required=True)\\n+ap.add_argument(\\\"--weights\\\", default=None, help=\\\"optional four comma-separated relevance weights\\\")\\n@@\\n shares /= shares.sum()\\n+weights = ([float(x) for x in a.weights.split(',')] if a.weights else [a.weight]*4)\\n+assert len(weights) == 4\\n@@\\n-    scores[q] = (qz + a.weight * robust_z(relevance[q])\\n+    scores[q] = (qz + weights[q] * robust_z(relevance[q])\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python - <<'PY'\\nimport json,numpy as np\\ns=np.load('/tmp/pool_scores.npz');q=s['quality'];o=np.load('/tmp/pool_offsets.npy')\\nfor x in ['ratio25','ratio30','ratio35','ratio40']:\\n a=json.load(open('/tmp/'+x+'.json'));tot=0;n=0;vals=[]\\n for i in a:\\n  if tot>=12_000_000:break\\n  tot+=o[i+1]-o[i];vals.append(q[i]);n+=1\\n print(x,n,np.quantile(vals,[0,.01,.1,.25,.5,.75,.9,1]),'late>',sum(i>=110000 for i in a[:n])/n,'eod estimate?')\\nPY\",workdir:\"/workspace\",yield_time_ms:10000,max_output_tokens:2000});\ntext(r.output);\n"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Update File: /workspace/build_candidate.py\\n@@\\n ap.add_argument(\\\"--dedup\\\", action=\\\"store_true\\\")\\n+ap.add_argument(\\\"--floor\\\", type=float, default=-2.0)\\n@@\\n-    if quality[i] < -2.0: continue\\n+    if quality[i] < a.floor: continue\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\n"
}

{
  "input": "const r = await tools.exec_command({cmd:\"python build_candidate.py --weight 3 --floor -1 --name r3floor1 && python build_candidate.py --weight 3 --floor 0 --name r3floor0 && python build_candidate.py --weight 3 --floor -3 --name r3floor3\",workdir:\"/workspace\",yield_time_ms:30000,max_output_tokens:2000});\ntext(r.output);\n"
}

{
  "input": "const r = await tools.write_stdin({session_id:56988,chars:\"\",yield_time_ms:30000,max_output_tokens:3000});\ntext(JSON.stringify(r));\n"
}

{
  "cell_id": "115",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}