Skip to main content
Managed SFT and DPO no longer accept new jobs. Your trained models stay deployable. To keep training, follow Migrate from Managed Training.
Managed Training ran supervised fine-tuning (SFT) and preference optimization (DPO and ORPO) jobs on Fireworks infrastructure from a dataset you uploaded, with no training code.

​
Existing jobs

Jobs you created before the pause stay in your account, and queued or running jobs finish normally. You can still:
  • View them on the Training page, with firectl sftj get and firectl sftj list (or firectl dpo-job get and firectl dpo-job list), or through the API.
  • Stop a running job: open the job and choose Actions → Stop, or run firectl sftj cancel <JOB_ID> (or firectl dpo-job cancel <JOB_ID>).
  • Delete a job from the same Actions menu, or with firectl sftj delete <JOB_ID> (or firectl dpo-job delete <JOB_ID>).
  • Export DPO metrics with firectl dpo-job export-metrics <JOB_ID>.
  • Deploy the model a job produced. See Deploying a trained model.
  • Re-run a finished job on the Training API: open the job and choose Actions → Re-run on Training API, or follow Migrate from Managed Training.
You can’t resume a failed or cancelled job. Re-run it on the Training API instead.

​
Training metrics

Per-step loss and learning-rate artifacts are retained for 7 days after they are written. After that, the job page explains that the chart is no longer available. If a job was configured with Weights & Biases, its metrics are also in your W&B project.

​
Methods

Supervised Fine-Tuning

Dataset format, job settings, and deployment for managed SFT jobs

Preference Optimization (DPO / ORPO)

Dataset format, monitoring, and deployment for managed DPO and ORPO jobs