scripts/job_he_full.sh

#!/bin/bash
# Claim 4 at FULL scale: all 164 HumanEval problems, baseline vs CCD.
# At n=32 CCD showed +9.38 with a unanimous 3-0 McNemar split, but 3 discordant
# pairs cannot reach p<0.05 (0.250 is the floor). n=164 should yield ~15
# discordant pairs -- enough to actually settle the paper's headline quality claim.
set -uo pipefail
pip install -q "transformers==4.46.2" "huggingface_hub<1.0" "datasets<4" "accelerate" 2>&1 | tail -1
python -c "
from huggingface_hub import snapshot_download
snapshot_download('ashishk1331/ccd-repro-code', repo_type='dataset', local_dir='/work')"
cd /work && mkdir -p outputs
nvidia-smi --query-gpu=name,memory.total --format=csv

python - <<'EOF' || exit 1
import torch, sys
if not torch.cuda.is_available(): print("ABORT: no CUDA"); sys.exit(1)
try:
    (torch.zeros(8,8,device="cuda",dtype=torch.bfloat16) @
     torch.zeros(8,8,device="cuda",dtype=torch.bfloat16)).cpu()
except Exception as e:
    print(f"ABORT: GPU unusable with torch {torch.__version__}: {e}"); sys.exit(1)
print(f"GPU OK: {torch.cuda.get_device_name(0)} torch {torch.__version__}")
EOF

python - <<'EOF'
from huggingface_hub import HfApi, snapshot_download
import glob, shutil
HfApi().create_repo('ashishk1331/ccd-repro-results', repo_type='dataset', exist_ok=True)
try:
    snapshot_download('ashishk1331/ccd-repro-results', repo_type='dataset', local_dir='/work/_prev')
    n = 0
    for f in glob.glob('/work/_prev/outputs/*.json'):
        shutil.copy(f, '/work/outputs/'); n += 1
    print(f'resumed {n} finished configs')
except Exception as e:
    print('no previous results:', e)
EOF

push () {
  python - <<'EOF' 2>&1 | tail -1 || true
from huggingface_hub import HfApi
HfApi().upload_folder(folder_path="outputs", path_in_repo="outputs",
                      repo_id="ashishk1331/ccd-repro-results", repo_type="dataset")
print("pushed")
EOF
}
run () {
  out="outputs/$1"; shift
  if [ -f "$out" ]; then echo "SKIP $out"; return; fi
  echo "=========== RUN $out : $* ==========="
  python scripts/run_eval.py "$@" --out "$out" || echo "!!!!! FAILED: $out"
  push
}

# full 164, the two arms that decide the paper's headline quality claim
run "c4full_he_baseline.json" --task humaneval --method baseline --limit 164 --temperature 0.0 --top-p 0.9
run "c4full_he_ccd.json"      --task humaneval --method ccd      --limit 164 --temperature 0.0 --top-p 0.9

echo "=================== HE FULL DONE ==================="
python - <<'EOF'
import json
b = json.load(open("outputs/c4full_he_baseline.json"))
c = json.load(open("outputs/c4full_he_ccd.json"))
import math
bs, cs = b["per_example_score"], c["per_example_score"]
b01 = sum(1 for x,y in zip(bs,cs) if x==0 and y==1)
b10 = sum(1 for x,y in zip(bs,cs) if x==1 and y==0)
n = b01+b10; k = min(b01,b10)
p = min(1.0, 2*sum(math.comb(n,i) for i in range(k+1))/2**n) if n else 1.0
print(f"baseline {b['score']:.2f}   CCD {c['score']:.2f}   delta {c['score']-b['score']:+.2f}")
print(f"McNemar: win {b01} / lose {b10}  ->  p = {p:.4f}   (paper claims +4.65)")
EOF
push