Product

Train when the data exists

Fine-tune is a job, not a six-week project. You send the dataset. You get back a versioned endpoint.

train · atlas-spSFT · LoRA

dataset.jsonl2.400 examples✓

loss 0.42

epoch

v1v2v3promoted

  • SFT or LoRA

    Gemma, or an adapter you already have. No rewrite of what the partner uses today.

  • The job goes up and comes down

    It runs on Cloud Run or GKE and stops when it is done. No cluster left on at month end.

  • Promote one, not all

    The new version goes to the partner who trained it. The others stay on the one they had.