Creating & Monitoring Jobs

~22 min · openai, jobs, hyperparameters, monitoring

Level 0Observer

0 XP0/43 lessons0/11 achievements

0/120 XP to next level120 XP to go0% complete

Create the job

Once your files are uploaded, kick off the fine-tuning job. "auto" on every hyperparameter is the right default for most projects — OpenAI tunes them based on dataset size.

Monitoring

The job goes through queued → running → succeeded/failed. Watch the events stream for training and validation loss; OpenAI emits a metrics line every ~50 steps.

Hyperparameter guide

Parameter	Default	When to override
`n_epochs`	auto (usually 3)	If validation loss rises early → reduce to 1 or 2.
`batch_size`	auto	Rarely needed. Auto picks a sane size based on dataset.
`learning_rate_multiplier`	auto	Lower (0.5–1.0) for small datasets, higher (1.5–2.0) for large.

Code

Create a job, then watch events + checkpoints·python

from openai import OpenAI
import time

client = OpenAI()

job = client.fine_tuning.jobs.create(
    training_file="file-abc123",
    validation_file="file-def456",
    model="gpt-4.1-mini-2025-04-14",
    suffix="my-custom-model",
    hyperparameters={
        "n_epochs": 3,
        "batch_size": "auto",
        "learning_rate_multiplier": "auto",
    },
    seed=42,  # reproducibility
)
print(f"Job ID: {job.id}, status: {job.status}")

# Poll for events as the job runs
last_seen = None
while True:
    events = client.fine_tuning.jobs.list_events(
        fine_tuning_job_id=job.id, limit=20
    )
    for e in reversed(events.data):
        if last_seen is None or e.created_at > last_seen:
            print(f"{e.created_at}: {e.message}")
    last_seen = events.data[0].created_at if events.data else last_seen

    j = client.fine_tuning.jobs.retrieve(job.id)
    if j.status in ("succeeded", "failed", "cancelled"):
        print(f"Final status: {j.status}")
        if j.status == "succeeded":
            print(f"Fine-tuned model: {j.fine_tuned_model}")
        break
    time.sleep(60)

# After completion: list checkpoints
checkpoints = client.fine_tuning.jobs.checkpoints.list(job.id)
for cp in checkpoints.data:
    print(f"step {cp.step_number}: {cp.fine_tuned_model_checkpoint}")
    print(f"  train loss: {cp.metrics.get('train_loss')}")
    print(f"  valid loss: {cp.metrics.get('valid_loss')}")

External links

Exercise

Using the file IDs from the previous lesson, kick off a fine-tuning job with seed=42 and three epochs. Poll the events stream until completion. Record training and validation loss at each checkpoint. Pick the checkpoint with the lowest validation loss as your production model.