Skip to main content
Training guide chapters All documentation

Model Training API

Fine-tune a separate customer model inside the FusionPointAI training SDK, using an agreed objective, prepared data and measurable evaluation.

Start here: the complete training journey

Training guide chapters

Commands and code stay in English in every language. Replace uppercase placeholders with values from your own account and paid job; do not paste real keys into shared notebooks or logs.

The training journey

The objective guides the work; actual optimization changes the student weights.
  1. Objective Define the task and acceptance criteria
  2. Teacher Prepare authorized reference material
  3. Student Train on the selected runtime
  4. Evaluation Compare with the baseline
  5. Delivery Export the independent model or adapter

Where the work happens

The API coordinates; the customer’s runtime performs weight training. No teacher weights are transferred.
  1. Your application API key, objective and approved job
  2. FusionPointAI API Session, reference access and prepared data
  3. Training runtime SDK, student model and GPU
  4. Local output Model package and evaluation report

From an objective to a separate model

Choose a provider plan and reference model, define the task, and configure the student model and runtime. TrainingProject can request authorized training material from the teacher, while the selected runtime loads the student and updates its weights. The result is a separate model or adapter, not a copy of teacher weights.

  • Start with a measurable objective, such as German document classification.
  • Keep teacher access, student model, runtime and evaluation criteria distinct.
  • TrainingClient remains available for existing paid job and session workflows.

Three parts, three responsibilities

Your API key authorizes service access; it is not a GPU. The API coordinates sessions, prepared data and reports. Your runtime supplies memory, storage and compute. The SDK performs local optimization and packages the output. Hosting may be provided by FusionPointAI, your data center, or a compatible notebook runtime.

  • Keep credentials out of model files and exported packages.
  • Review what objective and training material is sent to the teacher.
  • A completion report is not an independent certification of model quality.

Choose your reading path

New users can follow Quickstart, Concepts, Models, Datasets, Training and Evaluation in order. Integrators can go straight to the API and Python references. Operators should also read Recovery and Runtime deployment.

  • Review the plan before using compute or teacher requests.
  • Preserve an untouched baseline for comparison.
  • Export only after reviewing training and evaluation results.

API reference

EndpointPurpose and behavior
POST /api/v1/training/quotes Authorization: Bearer ai:training Validate eligible reference, plan and configuration; return a quote UUID, currency, fixed_total and expires_at. Use the approved plan and eligible student settings, not arbitrary client prices.
GET /api/v1/training/jobs Authorization: Bearer ai:training List the authenticated customer’s jobs. Inspect the returned collection and resource IDs rather than guessing numeric identifiers.
POST /api/v1/training/jobs Authorization: Bearer ai:training Create the account-owned job from a valid quote UUID. The response includes job ID, status, invoice ID and timestamps; creating a job does not prove payment.
GET /api/v1/training/jobs/{job} Authorization: Bearer ai:training Read one owned job using its job UUID, including live progress and delivery status.
GET /api/v1/training/jobs/{job}/artifacts/{file} Authorization: Bearer ai:training Download a published file belonging to this owned job. A local SDK ZIP is not automatically uploaded or available through this route.
POST /api/v1/training/sessions Authorization: Bearer ai:training Body: job UUID, runtime and optional positive reference_model_id. Return the new or reused session; reference changes are allowed only before preparation starts.
GET /api/v1/training/sessions/{session} Authorization: Bearer ai:training Read an owned session and its configuration, datasets, report and state. Reading state does not start training.
POST /api/v1/training/sessions/{session}/prepare Authorization: Bearer ai:training Send an empty JSON object to request a bounded preparation batch. Repeat while progress is made until ready, subject to the request budget.
GET /api/v1/training/sessions/{session}/datasets/{split} Authorization: Bearer ai:training Download train or validation as application/x-ndjson. Verify bytes, SHA-256 and rows against that session’s manifest.
POST /api/v1/training/sessions/{session}/events Authorization: Bearer ai:training Report real state, step, total_steps and optional loss/eval_loss metrics. Use only the supported state and error-code values.
POST /api/v1/training/sessions/{session}/complete Authorization: Bearer ai:training Submit artifact_name and the full real training report. The service checks method, base revision, weight changes, artifact digest and evaluation criteria.
POST /api/v1/training/sessions/{session}/cancel Authorization: Bearer ai:training Send an empty JSON object to cancel eligible service state. Stop the separate GPU process yourself when required; this is not pause/resume.
POST /api/v1/training/projects/authorize Authorization: Bearer ai:training POST /api/v1/training/projects/authorize with provider and positive reference_model_id. Revalidates the authenticated account’s paid training reference access and returns the owned job, grant, provider, reference model and expiry. It does not create an invoice or change a payment.
POST /api/v1/training/projects/teacher Authorization: Bearer ai:training POST /api/v1/training/projects/teacher with provider, reference_model_id, action (curriculum or lessons), objective, optional topic, count (1–4) and up to eight approved_examples (each at most 2,000 characters). Returns topics or lesson text/topic rows. Send only approved training material, never validation or final-test data.

FAQ

Can one objective start the automatic workflow?

It can initiate configuration. Actual training starts only after the required permissions, approved data, compatible student model, exact quote, payment and execution prerequisites are satisfied.

Do I receive the reference model’s original weights?

No automatic transfer is implied. The customer model is trained separately, and its deliverables and licence are defined in the approved quote.

Is performance identical to the reference model?

Not unconditionally. Agreed capability targets are assessed using documented benchmarks and the measured results are reported.

Must I purchase hosting?

No, when the approved deliverable is a downloadable artifact. Supported hosted deployments are priced separately from the one-time training project.

What happens to included API access afterwards?

It expires at the end of its recorded training period. You can then choose a separate available Unlimited API subscription.

Does one provider training plan cover more than one reference model?

Yes. One plan covers all available reference models from the subscribed provider. Choose a model for each training session, or switch to another available model before preparation, without another subscription. Access remains subject to the paid plan period and its expiry or earlier revocation.