Fine-Tune a Model#

Use the browser application to prepare training data and run model customization on connected NeMo Microservices. POC Factory coordinates the workflow; the remote platform supplies training and model-serving resources.

Prerequisites#

Before starting, confirm the application has a valid inference profile and connected NeMo Microservices endpoints. The backend also needs persistent, writable dataset and artifact storage. If endpoints are unavailable, ask the deployment operator to complete that connection first.

Run a Fine-Tune#

Start with a specific goal and review each plan and dataset before submitting training.

  1. Open Create → Model Finetuning → Dataset & Training.

  2. Enter a project name and goal. Choose Synthetic Dataset, Advanced Synthetic Dataset, or Upload Dataset. Uploaded JSONL must contain non-empty prompt and completion strings; preparation validates the file before training.

  3. Review the training-data plan and its evaluation and retry policies, then select Accept plan.

  4. Generate or prepare the dataset. Review sample records and accepted/rejected counts, then select Approve Dataset when prompted. Acceptance yield describes filtering efficiency, not model quality.

  5. Confirm the NeMo endpoints and choose a discovered Base model to finetune. Deployment-managed endpoint fields remain read-only.

  6. Select Continue to Fine-tuning and monitor the existing run through scheduling and training.

  7. When training completes, select Open Deployment & Evaluation, then Deploy Model. Wait for the deployment to report ready.

  8. Select Run Evaluation and review the scores, measured criteria, unavailable evidence, and recommendation.

Verify and Reuse the Result#

Training completion, model readiness, and evaluation are separate results. A ship recommendation means the recorded required numeric thresholds passed; narrative success criteria can remain unmeasured. Review that evidence before adopting the model.

An eligible Create retry-plan draft action prepares another plan for explicit review and acceptance; it does not automatically start another training run.

Use Download Dataset and Download Config Package to retain artifacts, and Use this fine-tune in a new POC to attach the run to generation. Select Clean Up Model when the deployed adapter is no longer needed, and review the reported cleanup results before assuming remote resources were removed.

Next Steps#

Use Troubleshooting for discovery, training, or evaluation failures. Follow Operate and restore to retain datasets and artifacts with application state.