Skip to main content
Training guide chapters All documentation

Teacher, student and training objective

Understand the difference between generating examples, changing weights and measuring a trained capability.

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.

Topic-by-topic training

Run within configured budgets and stop on failure or no progress.
  1. Baseline Measure existing task criteria
  2. Topic Select a prerequisite
  3. Practice Prepare lessons and update weights
  4. Assess Evaluate local held-out tasks
  5. Decide Advance, improve or stop

The teacher is a reference, not the deliverable

The teacher generates or reviews educational material through an authorized reference integration. The student is the trainable customer model loaded in the SDK. Generating an answer from a teacher does not update the student: a separate optimization step uses training batches to change student parameters. Teacher weights are not downloaded through an inference API.

  • One provider subscription covers the reference models included in that provider plan.
  • Selecting a different reference does not create another subscription.
  • Paid access period and job compatibility are checked before use.

Make the objective testable

“Train me in German” expresses a direction, but training needs a defined audience and measurable tasks. Specify whether you need conversation, translation, document understanding, writing, reasoning or a specialist domain. State acceptable error rates, privacy requirements and examples of unacceptable behavior. Keep hidden evaluation questions out of teacher-generated training material.

  • Example: classify German support messages into agreed categories.
  • Example: translate a held-out collection while preserving names, numbers and meaning.
  • Example: assess German writing with an agreed rubric rather than one subjective score.
Example · json 
{
  "objective": "Train a separate model to classify German support messages into agreed categories.",
  "language": "German",
  "performance_criteria": "Evaluate held-out messages, report per-class precision and recall, and review regression against the untouched baseline."
}

Use a bounded topic curriculum

TrainingProject can use explicit curriculum topics or ask the authorized teacher to propose a curriculum. It evaluates the untouched baseline, trains topic lessons within configured budgets, and stops on acceptance, budget exhaustion, no progress, cancellation or regression. The student capacity, data and compatible runtime determine the outcome.

  • Keep training, validation and final acceptance data separate.
  • Teacher sharing is restricted to approved training text; hidden evaluation data is never shared.
  • Reports retain topics completed and unmet criteria.