File size: 2,229 Bytes
cb7d6f0 2003020 cb7d6f0 2003020 cb7d6f0 | 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 | """Kaggle notebook body: build the director's training set, fine-tune, re-score.
Three subprocesses, same upgraded environment:
1. collect - sample the base model, keep every script that already satisfies the
TTS contract (acceptance is about 60% with the line format, so 160 scripts
costs roughly 35 minutes).
2. train - fine-tune the last six blocks and the head on those 160.
3. eval - score the tuned model on the same topic list as the baseline.
Everything runs out-of-process because transformers has to be upgraded to 5.x
and this notebook already imported huggingface_hub to fetch the bundle.
"""
import os
import shutil
import subprocess
import sys
BUNDLE = "/kaggle/working/webfm-eval"
ENV = dict(os.environ, WORK_ROOT="/kaggle/working")
def sh(cmd: str) -> None:
print(f"$ {cmd}", flush=True)
if subprocess.run(cmd, shell=True).returncode != 0:
raise SystemExit(f"failed: {cmd}")
def run(script: str, **env) -> int:
target = os.path.join(BUNDLE, "kaggle", script)
print(f"\n===== {script} {env} =====", flush=True)
rc = subprocess.run([sys.executable, target], env=dict(ENV, **env)).returncode
print(f"----- {script} rc={rc} -----", flush=True)
return rc
def main() -> None:
sh("pip install -q --upgrade 'transformers>=5' 'huggingface_hub>=1.31,<2' accelerate")
if not os.path.exists(os.path.join(BUNDLE, "kaggle", "director_probe.py")):
raise SystemExit("director_probe.py missing")
# 160 scripts collected on the first attempt (68% acceptance); reuse them
# instead of paying 25 minutes of sampling again.
seed = os.path.join(BUNDLE, "director", "sft.jsonl")
if os.path.exists(seed):
os.makedirs("/kaggle/working/director", exist_ok=True)
shutil.copyfile(seed, "/kaggle/working/director/sft.jsonl")
print("reusing collected scripts", flush=True)
else:
run("director_probe.py", MODE="collect", TARGET="160", MAX_TRIES="700")
if run("director_sft.py", EPOCHS="4", TRAIN_LAYERS="6") != 0:
raise SystemExit("training failed")
run("director_probe.py", MODE="eval", N_SAMPLES="12",
DIRECTOR_MODEL="/kaggle/working/director/model")
if __name__ == "__main__":
main()
|