Skip to main content
Continued fine-tuning starts a new fine-tune from a completed fine-tune. The new run trains on its own data with its own settings. You improve a trained model without retraining from scratch. For implementation details and code examples, see the Continue a fine-tuned model SDK guide.

How it works

A continued fine-tune is a new fine-tuning job that builds on a completed fine-tune, called the parent. SeekrFlow always starts the job from the parent’s latest saved checkpoint. The new run has its own training data, hyperparameters, and training method. The base model stays the same in every round. You can continue a model for any number of rounds. Each continued fine-tune records its parent. The lineage shows a model’s full history, from the run trained on the base model to the latest round, oldest first. Each round has a generation number, and generation 0 is the run trained from the base model. For example output, see View a fine-tune’s lineage in the SDK guide.

When to use continued fine-tuning

Continued fine-tuning provides value when:
  • Adding new data to a trained model without repeating earlier training
  • Refining a trained model with a different training method, such as reinforcement or preference tuning
  • Updating a model as your data changes over time
  • Building a model in stages, with each round saved as its own fine-tune

Parent model requirements

A fine-tune can serve as a parent when it meets these conditions:
  • It’s completed. Both full fine-tunes and LoRA fine-tunes qualify.
  • It belongs to your team.
  • It saved at least one checkpoint during training.

Continue with LoRA

A LoRA fine-tune produces a small add-on called an adapter instead of a full set of model weights. You can continue a LoRA fine-tune with LoRA, and you can continue a full fine-tune with LoRA. SeekrFlow decides what the new run produces.
  • When you continue a LoRA fine-tune with LoRA, SeekrFlow combines the new adapter with the parent’s into one adapter. The combined adapter’s rank is the sum of the two. It serves on the base model’s shared deployment, like any other LoRA adapter. If the combined adapter is too large for the shared deployment, SeekrFlow merges it into full model weights instead.
  • When you continue a full fine-tune with LoRA, SeekrFlow merges the new adapter into the full model weights.
  • Full model weights deploy as their own model. Every later round in that lineage also produces full model weights.
A LoRA fine-tune continues only with LoRA. Continued fine-tuning supports plain LoRA. For the supported settings, see Continue with LoRA in the SDK guide.

Supported training methods

You can continue a full fine-tune with any of these methods:
Last modified on November 16, 2026