Fine-tuning on data you can account for.
We train smaller, faster models on your data, inside your environment, and keep a record of exactly which data produced which model.
What you get
A model, and the evidence behind it.
A cleaned dataset
Duplicates removed, quality filtered, and personal data detected and redacted, including local identifiers such as NRIC numbers and IBANs. A person reviews a sample.
Training runs
LoRA or QLoRA for most work, preference tuning where tone and judgement matter, and distillation when a small model must replace a large one.
Lineage
Every dataset and model version is hashed and recorded, with pinned code and settings, so any model can be rebuilt from the same data, code and configuration, with results verified against the recorded evaluation.
Memorisation tests
Planted canary strings and extraction probes check whether the model repeats training data before it is released.
Before-and-after report
Results on your tasks for the base model and the tuned one, with any regressions called out.
An erasure plan
An inventory of what went into training, so a model can be retrained without a record when someone asks to be erased.
How it works
Five steps, each one recorded.
01
Inventory
Sources, purposes and owners of the data. You decide the lawful basis; we supply the facts for your impact assessment.
02
Clean and split
Redaction, de-duplication and a test set kept apart, with checks that none of it leaks into training.
03
Train
Small runs first. We choose the method from results, not habit.
04
Test
Task accuracy, memorisation and safety, against the base model.
05
Release
A training card for the run, and sign-off from your owner before it goes live.
Technical detail
Methods and tools.
Methods
Supervised fine-tuning with LoRA or QLoRA, preference tuning, and distillation with the teacher model’s licence checked.
SFT · LoRA · QLoRA · DPO · ORPO · KTO
Lineage
Hashed artefacts, pinned containers and a training card for every run.
DVC · lakeFS · MLflow
Privacy
Detection with custom recognisers, pseudonymisation, and differential privacy where the accuracy cost is accepted.
Presidio · pseudonymisation · DP
Synthetic data
Generated examples to cover rare cases, or to train where real records can’t be used. Each set is checked for leaked personal data and tested against real data held out from training.
seeded generation · privacy checks · real-data test set
Regulation
Built to support your obligations.
| Framework | What it asks | How we support it |
|---|---|---|
| GDPR Art. 5(1)(b), 6(4) | Further use of personal data must be compatible with its original purpose. | We record sources and purposes. You decide whether training is compatible. |
| GDPR Art. 9 | Extra protection for health and other special-category data. | Health data is minimised and redacted before training, and access is logged. |
| GDPR Art. 35 | A data protection impact assessment for high-risk processing. | We supply the technical inputs for your assessment. |
| EDPB Opinion 28/2024 | A trained model is not automatically anonymous. | We run extraction tests and record the results. |
| Singapore PDPA and PDPC AI guidelines | Accountability for personal data used in AI systems. | Data inventory and purpose records you can show. |
| EU AI Act Art. 10 | Data governance for high-risk systems. | Dataset documentation built in from the start. |
Questions
What risk teams ask.
Does our data leave our environment?
Not in the your-cloud, data-centre or air-gapped options. In the hosted option, data stays in a single-tenant environment we run for you, in the region you choose. Your data is never used to train anything for anyone else. This is a contract term.
Do your engineers see raw personal data?
They review samples to check quality. That happens inside your environment, under your access controls, and every session is logged.
How do you handle erasure requests?
We keep an inventory of training data, so we can remove the record and retrain the adapter from the cleaned dataset.
Can you prove the test set wasn’t in training?
Yes. The split is fixed and hashed before training, and we run overlap checks you can repeat.
What we don’t promise
Some claims we won’t make.
- Redaction that catches every piece of personal data. We measure it and review samples.
- That a trained model is anonymous. That’s decided case by case, and we give you the test results.
- That no person ever sees the data. Engineers review samples, under your controls.
A smaller model that knows your work.
Tell us the task and the data. We’ll tell you whether fine-tuning is worth it.