| #!/bin/bash |
| |
| |
| 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 |
|
|
| python scripts/run_eval.py --task trip --method ccd_ds --limit 3 --out outputs/diag_ds.json |
|
|
| python - <<'EOF' |
| import json |
| import numpy as np |
| r = json.load(open("outputs/diag_ds.json")) |
| print("\n=============== Eq. (20) BUDGET DIAGNOSTICS ===============") |
| print(f"mean steps {r['mean_steps']:.1f} (uniform baseline = 256) speedup {r['speedup_vs_uniform']:.2f}x") |
| print(f"fallback steps (|I^c_t| = 0): {r['fallback_steps']}") |
| ic = np.array(sum(r["ic_sizes"], [])) |
| st = np.array(sum(r["n_stable"], [])) |
| bud = np.array(sum(r["budgets"], [])) |
| print(f"\n|I^c_t| : mean={ic.mean():.2f} hist={np.bincount(ic, minlength=5)[:5]} (V=4)") |
| print(f"n_stable : mean={st.mean():.2f} hist={np.bincount(st, minlength=5)[:5]}") |
| print(f"budget : mean={bud.mean():.2f} hist={np.bincount(bud, minlength=5)[:5]}") |
| nz = ic > 0 |
| print(f"\nAmong steps with |I^c_t|>0: mean |I^c_t| = {ic[nz].mean():.2f}, " |
| f"mean stable = {st[nz].mean():.2f}, stable fraction = {st[nz].sum()/ic[nz].sum():.3f}") |
| print("\nIf |I^c_t| is large but n_stable ~ 0, the argmax-stability heuristic is the") |
| print("bottleneck. If |I^c_t| itself is ~1, the top-V intersection is the bottleneck.") |
| print("===========================================================") |
| EOF |
|
|