Datasets:
File size: 5,129 Bytes
46e27b5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 | #!/usr/bin/env python3
"""Bakeoff orchestrator: train + eval every BIO arm with identical settings.
One arm = one subprocess = one variable. Arms never share state. Any arm that
crashes is recorded as failed and the run continues, so one bad candidate
cannot abort the bakeoff.
Usage: run_bakeoff.py [--only tag1,tag2] [--skip-eval]
"""
import argparse
import json
import os
import subprocess
import sys
import time
SCRIPTS = "/opt/vox/sandbox/scripts"
PY = "/opt/vox/venv/bin/python"
CKPT = "/opt/vox/temp/ckpt"
BENCH = "/opt/vox/sandbox/benchmarks"
# Accuracy-first ordering. ONNX/casing/latency are NOT selection filters.
ARMS = [
("distilbert-cased", "distilbert/distilbert-base-cased"),
("bert-base-cased", "google-bert/bert-base-cased"),
("roberta-base", "FacebookAI/roberta-base"),
("deberta-v3-base", "microsoft/deberta-v3-base"),
("modernbert-base", "answerdotai/ModernBERT-base"),
("electra-base-disc", "google/electra-base-discriminator"),
]
COMMON = "--seed 42 --epochs 8 --batch-size 16 --lr-encoder 3e-5 --lr-head 1e-3 --max-len 128"
def sh(cmd, log):
with open(log, "a") as f:
f.write(f"\n$ {' '.join(cmd)}\n")
f.flush()
p = subprocess.run(cmd, stdout=f, stderr=subprocess.STDOUT)
return p.returncode
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--only", default="")
ap.add_argument("--skip-eval", action="store_true")
args = ap.parse_args()
only = {x for x in args.only.split(",") if x}
arms = [a for a in ARMS if not only or a[0] in only]
os.makedirs(CKPT, exist_ok=True)
os.makedirs(BENCH, exist_ok=True)
os.makedirs("/opt/vox/temp/logs", exist_ok=True)
summary = []
for tag, hf in arms:
out_dir = os.path.join(CKPT, tag)
log = f"/opt/vox/temp/logs/arm_{tag}.log"
t0 = time.time()
print(f"=== {tag} ({hf})", flush=True)
if not os.path.exists(os.path.join(out_dir, "train_meta.json")):
rc = sh([PY, "-u", f"{SCRIPTS}/train_span.py", "--model", hf,
"--out", out_dir] + COMMON.split(), log)
if rc != 0:
print(f" TRAIN FAILED rc={rc} -> see {log}", flush=True)
summary.append({"tag": tag, "model": hf, "status": "train_failed"})
continue
else:
print(" checkpoint exists, skipping train", flush=True)
meta = json.load(open(os.path.join(out_dir, "train_meta.json")))
print(f" trained {meta['n_train']} ex, {meta['params_m']}M, "
f"loss {meta['train_log'][0]['train_loss']}->{meta['train_log'][-1]['train_loss']}, "
f"{meta['wall_s']}s", flush=True)
if args.skip_eval:
continue
ej = os.path.join(BENCH, f"ft_{tag}.json")
rc = sh([PY, "-u", f"{SCRIPTS}/eval_span.py", "--model", out_dir, "--tag", f"ft_{tag}",
"--out", ej, "--cpu-latency", "--export-onnx", "--onnx-int8"], log)
if rc != 0 or not os.path.exists(ej):
print(f" EVAL FAILED rc={rc} -> see {log}", flush=True)
summary.append({"tag": tag, "model": hf, "status": "eval_failed"})
continue
r = json.load(open(ej))
b = r["best"]
summary.append({
"tag": tag, "model": hf, "params_m": r["params_m"],
"exact_match": b["exact_match"], "span_word_f1": b["span_word"]["f1"],
"precision": b["span_word"]["precision"], "recall": b["span_word"]["recall"],
"over_deletion_rate": b["over_deletion_rate"],
"harmful_span_rate": b["harmful_span_rate"],
"thr": b["threshold"], "passes_gate": b["passes_gate"],
"clean_exact_match": b["per_category"].get("CLEAN", {}).get("exact_match"),
"dis_exact_match": b["per_category"].get("DISFLUENCY", {}).get("exact_match"),
"cpu_p95_ms": r.get("cpu_latency", {}).get("p95_ms"),
"onnx_parity_rate": r.get("onnx", {}).get("parity_rate"),
"int8_compression": r.get("onnx", {}).get("int8_compression"),
})
print(f" EM={b['exact_match']} f1={b['span_word']['f1']} "
f"overdel={b['over_deletion_rate']}% harmful={b['harmful_span_rate']}% "
f"gate={b['passes_gate']} cpu_p95={r.get('cpu_latency',{}).get('p95_ms')}ms "
f"parity={r.get('onnx',{}).get('parity_rate')}%", flush=True)
with open("/opt/vox/temp/logs/bakeoff_summary.json", "w") as f:
json.dump(summary, f, indent=2)
print("\n=== BAKEOFF SUMMARY ===", flush=True)
for s in sorted([x for x in summary if "exact_match" in x],
key=lambda x: -x["exact_match"]):
print(f" {s['tag']:20} EM={s['exact_match']:6} f1={s['span_word_f1']:.3f} "
f"P={s['precision']:.3f} R={s['recall']:.3f} overdel={s['over_deletion_rate']}% "
f"harm={s['harmful_span_rate']}% gate={s['passes_gate']} "
f"cpu_p95={s['cpu_p95_ms']}ms", flush=True)
print("artifacts: /opt/vox/sandbox/benchmarks/ft_<tag>.json", flush=True)
if __name__ == "__main__":
main() |