scripts/hf_jobs_train_syllabus.py

# /// script
# requires-python = ">=3.10"
# dependencies = [
#   "huggingface_hub>=0.26.0",
# ]
# ///
"""Run syllabus LoRA SFT on Hugging Face Jobs (GPU).

Downloads finetune JSONL from the Hub, clones training-pipeline, runs readiness
gates, trains with OOM-safe defaults, and pushes the adapter to a Hub model repo.

Override via environment variables:
  HF_DATASET_REPO   (default: Dev-the-dev91/syllabus-finetune)
  HF_OUTPUT_MODEL   (default: Dev-the-dev91/syllabus-extractor-lora)
  HF_PIPELINE_REPO  (default: https://github.com/madch3m/training-pipeline.git)
  HF_MAX_LENGTH     (default: 2048)
  HF_TRAIN_EPOCHS   (default: 3)
  HF_JOB_FLAVOR     (informational only when run locally)
"""

from __future__ import annotations

import os
import subprocess
from pathlib import Path

from huggingface_hub import HfApi, hf_hub_download

HF_USER = os.environ.get("HF_USER", "Dev-the-dev91")
DATASET_REPO = os.environ.get("HF_DATASET_REPO", f"{HF_USER}/syllabus-finetune")
OUTPUT_MODEL = os.environ.get("HF_OUTPUT_MODEL", f"{HF_USER}/syllabus-extractor-lora")
PIPELINE_GIT = os.environ.get(
    "HF_PIPELINE_REPO",
    "https://github.com/madch3m/training-pipeline.git",
)
MAX_LENGTH = os.environ.get("HF_MAX_LENGTH", "2048")
NUM_EPOCHS = os.environ.get("HF_TRAIN_EPOCHS", "3")
WORK = Path(os.environ.get("HF_WORK_DIR", "/tmp/training_pipeline"))


def run(cmd: list[str], *, cwd: Path | None = None) -> None:
    print("+", " ".join(cmd), flush=True)
    subprocess.run(cmd, check=True, cwd=cwd)


def uv_run(script_args: list[str], *, cwd: Path) -> None:
    """Run a repo script using the synced project venv (--extra train)."""
    run(["uv", "run", "--extra", "train", *script_args], cwd=cwd)


def _require_hub_token() -> str:
    token = os.environ.get("HF_TOKEN", "").strip()
    if not token or token in ("$HF_TOKEN", "${HF_TOKEN}"):
        raise SystemExit(
            "HF_TOKEN is missing or still the literal '$HF_TOKEN' placeholder.\n"
            "Resubmit with: uv run python scripts/submit_hf_training_job.py\n"
            "Or CLI: hf jobs uv run --secrets HF_TOKEN <script-url>\n"
            "(Do not pass secrets={'HF_TOKEN': '$HF_TOKEN'} from Python — use the real token.)"
        )
    return token


def main() -> None:
    token = _require_hub_token()
    api = HfApi(token=token)
    who = api.whoami()["name"]
    print(f"Hub user: {who}")
    print(f"Dataset: {DATASET_REPO}")
    print(f"Output model: {OUTPUT_MODEL}")

    if WORK.exists():
        run(["rm", "-rf", str(WORK)])
    run(["git", "clone", "--depth", "1", PIPELINE_GIT, str(WORK)])
    # Create .venv with train extras so `uv run --extra train` sees datasets/trl/torch.
    run(["uv", "sync", "--extra", "train"], cwd=WORK)

    data_dir = WORK / "data" / "finetune"
    data_dir.mkdir(parents=True, exist_ok=True)
    for name in ("train.jsonl", "valid.jsonl"):
        cached = hf_hub_download(
            repo_id=DATASET_REPO,
            filename=name,
            repo_type="dataset",
        )
        (data_dir / name).write_bytes(Path(cached).read_bytes())
        print(f"Fetched {name} from {DATASET_REPO}")

    train_path = data_dir / "train.jsonl"
    valid_path = data_dir / "valid.jsonl"

    uv_run(
        [
            "validate_training_readiness.py",
            "--train-jsonl",
            str(train_path),
            "--valid-jsonl",
            str(valid_path),
            "--strict",
        ],
        cwd=WORK,
    )

    out_dir = WORK / "artifacts" / "hf_syllabus_extractor"
    uv_run(
        [
            "train_hf_structured_extractor.py",
            "--train-jsonl",
            str(train_path),
            "--valid-jsonl",
            str(valid_path),
            "--model-name",
            "Qwen/Qwen2.5-0.5B-Instruct",
            "--output-dir",
            str(out_dir),
            "--max-length",
            MAX_LENGTH,
            "--per-device-train-batch-size",
            "1",
            "--gradient-accumulation-steps",
            "8",
            "--num-train-epochs",
            NUM_EPOCHS,
            "--bf16",
            "--disable-mlflow",
        ],
        cwd=WORK,
    )

    api.create_repo(OUTPUT_MODEL, repo_type="model", exist_ok=True)
    api.upload_folder(
        folder_path=str(out_dir),
        repo_id=OUTPUT_MODEL,
        repo_type="model",
        commit_message="LoRA adapter from HF Jobs syllabus SFT",
    )
    print(f"Uploaded adapter: https://huggingface.co/{OUTPUT_MODEL}")


if __name__ == "__main__":
    main()