Private and sensitive data
Fine-tune LLMs on private and sensitive data
Fine-tune open models on medical, legal and financial records. See what Tensorant stores, where your data and GPUs run, and what stays in your cloud.
The problem
Why fine-tune on sensitive records instead of sending them to an API model?
Every request to a general API model sends the record in it, such as a patient note, a contract or a bank statement, to that model's provider.
You choose where a fine-tuned open model is served, and you keep its adapter. You can publish it on Tensorant, or download the adapter and run it with its base model on your own hardware.
How it works
What fine-tuning changes
- The model learns your format and your rules from examples, so your prompts can carry less context, such as fewer instructions and examples.
- You can score it on held-out records before you publish it. An experiment scores the base model, trains an adapter and scores the adapter on the same test examples, which the version holds out of training.
In practice
What Tensorant stores, and where
- The content of your sources, examples and versions is kept in your organization's storage: Tensorant storage by default, or your own storage. Tensorant's database keeps records such as your account, project settings, file names, review states and scores, with references to that content. The content passes through Tensorant's servers when they work on it, such as when you import, generate, review or score examples. See What runs where.
- Your files, trained weights and GPU machines can stay in your own cloud, with your own S3 bucket, which every organization can use for free, and your own GPU account, which comes with the Own compute subscription. Training can run in your account on RunPod, Verda, Lambda, AWS, Azure or Google Cloud. Test sessions, experiments and published models in your own account run only on RunPod, with your RunPod volume as storage. Requests to a published model still pass through Tensorant on their way to its GPU. See Choose a setup guide, Connect S3 storage and What runs where.
- By default, Tensorant keeps your files in Tensorant storage, separately for each organization, and runs your training runs and published models on GPU machines it provides. See Where your data lives.
- You can download a trained adapter as a ZIP, and each download is recorded in your organization's audit trail. See The adapter.
- If a recipe generates examples with a model endpoint, that endpoint's provider sees the full text of each passage. See what is sent to the endpoint. You can generate with a model you deploy in Tensorant instead. An endpoint's provider also sees each example a quality judge checks and each prompt an evaluation sends it. An evaluation's LLM judge must be a model endpoint, and its provider receives each answer it scores, with the prompt, the reference answer and the full source text of the test example.
- Review flags text that looks like an email address or a phone number, and can replace it with
[EMAIL]or[PHONE]in the example. It does not detect names, addresses or other personal data, and it leaves your source files and existing versions unchanged. See Redact personal data.
Models
Models to start with
Open models whose weights you can run where you choose. Every family Tensorant trains is in Supported models.
Qwen/Qwen3-4B-Instruct-2507
A 4B instruct model that answers directly, without a reasoning block. Try it first.
mistralai/Mistral-7B-Instruct-v0.3
A 7B instruct model to try when a 4B model falls short on your test set.
unsloth/Llama-3.2-1B-Instruct
An ungated copy of Llama 3.2 1B Instruct, the smallest of these three.
Steps
How to do it in Tensorant
Choose where data and GPUs live
Before your first project, choose Tensorant storage or your own storage: your own storage can be chosen only while the organization has no projects. GPUs come from Tensorant compute or your own GPU account, chosen on each run. See Tensorant compute and storage.
Import your records
Upload your examples as CSV or JSON Lines, with question and answer columns or as conversations, or import a Hugging Face dataset. See Sources.
Review and redact
Approve examples in Review, and replace email addresses and phone numbers where you need to. Check for names and other personal data yourself. See Redact personal data.
Freeze a version
Create a version. It splits the approved examples into training, validation and test examples, and holds the test examples out for scoring.
Run an experiment
An experiment scores the base model, trains an adapter and scores the adapter on the same test examples.
Download or publish the adapter
Download the adapter from its run. See The adapter. Or give it a name in API models and call it through the API.
Plans
What you need
- Text fine-tuning, images, evaluations, publishing and adapter downloads are on every organization, which pays as it goes from credits. Each training run pays a training-token fee. See Training-token fee.
- GPUs come from Tensorant compute, charged per second from prepaid credits, which an owner adds under Settings, Plan and billing (the first payment turns Tensorant compute on), or from a GPU account you connect with the Own compute subscription. See connections.
Fine-tuning
Fine-tuning vs RAG: build a grounded support assistant
Choose retrieval, fine-tuning, or both for a support assistant, with source-conditioned examples, held-out evaluation, and Tensorant API steps.
Evaluation
How to prevent data leakage in LLM evaluation
Prepare held-out LLM evals in Tensorant, separate related sources and prompts, review near-duplicates, and preserve trustworthy tests as your dataset grows.
Evaluation
How to evaluate a fine-tuned LLM against its base model
Build useful LLM evals in Tensorant, score a support-ticket classifier, compare base and tuned models, and investigate regressions before deployment.
Classification and routing
Answer with one label from your list: intents, ticket routes, what needs a person.
JSON extraction
Pull the fields your code needs out of invoices, forms and emails, as JSON that fits your schema.
Replace a costly LLM API
Train an open model on the answers your API model gives, and switch when it scores as well.
Small language models
Fine-tune a 1B to 4B model for one narrow job, and download the adapter.
Scoring and ranking
Score each candidate answer with a calibrated probability, with a decision model such as kev.
Reinforcement learning
Train with GRPO against a reward that checks each answer: a label, a number, JSON or your own code.
Migrate from OpenAI fine-tuning
OpenAI ends fine-tuning in January 2027. Move your data, train an open model and compare the two.
Start with one dataset.
Create an account and import your data, or book a demo and we will show you the product.