Skip to main content
This guide connects the full workflow: collect or upload data, shape it into datasets, run evals, launch training, and move the result toward deployment.
Use this page as the hub for onboarding flows that point to fine-tuning, evals, deployment, and demo projects.

Get started with fine-tuning

Start from real inference data and launch a training run.

Get started with evals

Build a repeatable quality loop before or after training.

E2E fine-tuning and deployment

Connect datasets, evals, training, and deployment into one workflow.

Demo project

Start with a concrete example before you wire up production data.

​
Get Started with fine-tuning

  1. Collect or upload data. Capture production traffic with the dashboard integration guide, or upload historical JSONL logs through /fine-tuning/dataset-uploads.
  2. Create a training dataset. Use the Datasets surface in the dashboard, or inspect exported datasets with /cli/datasets.
  3. Check your success criteria. If you already know how you want to score the model, define or review evals first so you can compare checkpoints later.
  4. Launch training. Start a training run from the dashboard’s Training Jobs area or monitor existing runs with /cli/training.
  5. Review results before rollout. Compare outputs, inspect logs, and validate the model against the evals that matter for your task.
InputWhy it mattersWhere to start
Production tracesBest match for real usageDashboard integration guide
Historical logsGood if you already have request/response pairs/fine-tuning/dataset-uploads
Filtered dataset exportBest for repeatable training runs/cli/datasets

​
Get Started with evals

  1. Choose representative inputs. Start from the same production traces or uploaded logs you plan to optimize against.
  2. Create an eval dataset. Save filtered inferences as an eval dataset in the dashboard, or review datasets via /cli/evals.
  3. Define your rubric. Decide whether you want task-specific scoring, judge-model scoring, or a lightweight pass/fail check.
  4. Run the eval. Compare the current baseline model, a candidate model, or a newly trained checkpoint.
  5. Loop the result back into training. Use failures and low-score samples to tighten the next dataset revision.

​
When to run evals

  • Before training, to establish the current baseline.
  • During training, to compare checkpoints and catch regressions early.
  • After training, to decide whether the model is ready for rollout.

​
E2E fine-tuning and deployment

  1. Collect or upload request data. Use the dashboard integration guide for live traffic or /fine-tuning/dataset-uploads for historical logs.
  2. Build both training and eval datasets. Training data teaches the model; eval data tells you whether it improved.
  3. Launch training. Track runs and logs from the dashboard or via /cli/training.
  4. Run evals against the resulting checkpoint. Use the same rubrics and datasets you trust for production decisions.
  5. Deploy the winning model. Once quality, latency, and cost look right, move the model toward your production path.
  6. Keep observing. Route production traffic back through Inference.net so new traces continue improving the next iteration.

​
A practical rule of thumb

If a model is not yet good enough, improve the dataset and eval design first. If the model is good enough but too slow or expensive, push the workflow toward distillation, deployment, and tighter production monitoring.

​
Get started with a demo project

If you want a concrete example before using your own production data, start with a task that already has strong docs and sample material.

Schematron extraction notebook

Walk through a real extraction workflow using Schematron and a typed schema.

Quickstart

Make a direct API call before you wire up larger workflows.

Schematron docs

Learn the model surface, expected inputs, and recommended usage pattern.

Dataset uploads

Bring your own JSONL logs once you are ready to move beyond a demo.

​
What to do after the demo

  • Replace sample inputs with a small slice of your own production or historical data.
  • Save the best examples into a reusable dataset.
  • Add an eval loop before you scale the workflow up.
  • Move the successful workflow into the dashboard or CLI surfaces that fit your team.