Skip to content

Product tour

How fine-tuning works on Tensorant

Five screens from the console, on a sample project: review the data, train, evaluate, serve, and run every step as one experiment.

Watch the product demo

The five steps

  1. Review the data before you train on it.

    Upload documents and spreadsheets, or import a Hugging Face dataset. Approve, edit or reject each example.

    • Quality checks flag missing evidence, repetition and personal data.
    • Starter recipes write examples from your documents, each quoting its source.
    • Frozen versions keep training and test examples apart.
  2. Fine-tune without writing a training script.

    Pick a base model, a method (SFT, DPO or RL), LoRA or QLoRA, and 1 to 8 GPUs. RL trains with LoRA on one GPU. Tensorant checks the model fits before anything starts.

    • Every run has a time limit and an hourly price cap.
    • Follow the loss, the time and the cost while it trains.
    • The GPU is released when the run ends or is cancelled.
  3. See whether fine-tuning helped.

    The base and tuned models are scored on the same test examples with the same settings, answer by answer.

    • Presets for Q&A, classification, JSON and numeric answers.
    • An optional LLM judge scores answers against your rubric.
    • See which answers got better and which got worse.
  4. Deploy it behind an OpenAI-compatible API.

    Publish the model under a name and call it with the OpenAI SDK. Keep it always on, or let it sleep and wake on the next request.

    • Ask the base and tuned models the same question side by side.
    • A sleeping model costs no GPU time.
    • Keys per application, each with its own limits.
  5. Run every step as one experiment.

    Baseline, training, deployment, comparison and cleanup run in order once you approve the cost.

    • Set the score the tuned model has to reach.
    • See the score change and how certain it is.
    • Test deployments stop on their own.
app.tensorant.ioSample project
The Review screen: a queue of pending examples, one example's conversation with Approve and Reject, and its source and quality checks beside it.The Review screen: a queue of pending examples, one example's conversation with Approve and Reject, and its source and quality checks beside it.
Each example beside its source and its quality checks.

What it replaces

  • chunk_pdfs.py
  • review_examples.xlsx
  • train.py
  • eval_notebook.ipynb
  • serve.sh
  • # TODO: stop the GPU

Go deeper

Each stage has its own page

What a stage does, how it is set up and where its docs are.

Start with one dataset.

Create an account and import your data, or book a demo and we will show you the product.