Fine-tuning and distillation

Nuvora prepares and governs customization; your own trainer does the GPU work.

Datasets

POST /api/datasets {name, content} stores a JSONL dataset after validating it. Send dry_run: true to check it without storing.

Format	Each line
chat	{"messages": [{"role": …, "content": …}, …]}
completion	{"prompt": …, "completion": …}
prompts	{"prompt": …}, used for distillation

A dataset needs 10–100,000 records and at most 20 MiB. Every record passes the workspace guardrails, so blocked content never reaches a trainer. Listing datasets never returns their content.

Recipes

A recipe names a base model, a method and a dataset_id:

Method	What happens
lora, qlora	The dataset goes to the trainer as is, with rank and epochs
distillation	A teacher_model answers every prompt first, then the student is trained on those answers

POST /api/recipes/{id}/run starts the job. Without a trainer, POST /api/recipes/{id}/export gives you a TrainingRecipe to run elsewhere.

Connect a trainer

export NUVORA_TRAINER_URL='https://trainer.internal'
export NUVORA_SECRET_TRAINER_TOKEN='…'
# when the trainer doesn't report where the model is served:
export NUVORA_TRAINER_SERVING_URL='https://serving.internal/v1'
export NUVORA_TRAINER_SERVING_KEY_ENV='NUVORA_SECRET_SERVING_KEY'

The trainer's host must be on NUVORA_PROVIDER_HOSTS. Nuvora calls:

• POST /v1/training/jobs with base_model, method, hyperparameters, dataset (format, records, digest, content) and suffix; it expects an id back.
• GET /v1/training/jobs/{id} until status is succeeded, completed or success (or failed, error, cancelled).

While training, the job waits as waiting_external, for up to seven days. When it succeeds, Nuvora registers the result (result.model or fine_tuned_model) as a new model you can evaluate before anyone uses it.