Skip to main content

Fine-tuning Otto

Tune a custom Otto model on your tenant's example conversations to improve response style, terminology, and escalation behaviour.

When to fine-tune​

Fine-tune when you want Otto to learn:

  • Organization-specific terminology (internal service names, severity labels, team handles)
  • Escalation patterns specific to your IT, HR, or finance workflows
  • House response style (tone, formatting conventions)
  • Structured-action outputs aligned to your downstream system schemas

If your need is "make Otto know about article X", update your knowledge base or connect a new source - don't fine-tune. Fine-tuning shapes behaviour; retrieval supplies facts.

Prepare training data​

Training data is a JSONL file. Each line is one conversation example:

{"messages": [{"role": "system", "content": "..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}

Quality over quantity: 50-200 high-quality examples typically beat 1,000 noisy ones.

Start training​

Tune a Rezolve Otto model dialog with Display name, Identifier, Description, System prompt, and Training data (JSONL) fields and Cancel / Start training buttons

  1. Open Settings → Models and click Tune model.
  2. Fill Display name, Identifier, optional Description and System prompt.
  3. Upload your JSONL file (max 100 MB).
  4. Click Start training. Training typically takes 10-20 minutes.

After training​

  • The custom model appears in the agent model picker like any other variant.
  • Test on a few example prompts in the Testing tab before switching production agents over.
  • You can train multiple iterations side by side and A/B compare.

Tenant isolation and erasure​

  • Training data is stored in tenant-isolated storage with the customer's data residency and retention policies.
  • Trained model weights live only in the tenant's namespace; never shared across tenants and never folded back into the base Otto family.
  • On customer request (right-to-erasure, contract termination, data-residency change), training data and trained weights are deletable within the contractual SLA.

Pricing​

Fine-tuning is included in the Enterprise tier. See your account team for training quotas.