Data
Your data fits your model. Nothing else.
You bring a corpus and a written standard. We use them to fit and evaluate your private model. Your data never trains the shared base or another customer's fine-tune, and the tailored weights that result are your property.
A corpus and a standard
You need two things: examples of work done well, and a written note of the bar you hold. No labelling pipeline, no schema, no ML team required.
Examples of the work done well
Documents, conversations, decisions and outputs that show your company at its best. There is no required format and no pipeline to build before you bring them.
The bar you hold
A plain-language note of what good looks like: tone, house rules, what the model should and should not do. One document, written once, used to align and evaluate.
Fit and eval only
Your data has two jobs: it fits the model and it sets the bar. That is all.
Fine-tuning your model
Your corpus is used to fine-tune a private model on the way your company writes, decides and communicates. The training run is isolated: no other customer's data is in the same run.
Scoring against your standard
A held-out portion of your examples is used to build the eval that scores every version of the model. Your standard is the bar. Nothing else sets it.
Not the shelf, not anyone else
Your data is never used to train the shared base model and never used to train another customer's fine-tune. That rule is contractual, not a default that can be changed by a setting.
The weights are yours, in writing
The tailored weights are your property. That is stated in your contract and does not change if you cancel, downgrade or move on.
Export at any time
You may export your tailored weights in a standard format at any point during or after an engagement. The export is provided in a format compatible with common open-weight runtimes so you can continue to use what you built without depending on Kethra to serve it.
Data residency
Training data and weights are stored by default in the United States. EU and UK data residency are available on request for all paid tiers. Specific-region requirements can be accommodated within the Enterprise configuration, where all data and training may remain inside your own perimeter.
Your data is not the shelf
Kethra operates a strict separation between customer training data and any shared model infrastructure. Your corpus does not improve the base model that Kethra starts from, and it does not improve any other customer's fine-tune. The improvements made with your data are captured in your weights and nowhere else.
Your data on your schedule
What we retain and for how long
Raw training data is retained for the duration of the active engagement unless you request earlier deletion. We retain a minimal record of the training run, including hyperparameters and eval results, for the purposes of auditability and model refinement. This record does not include your raw corpus content.
After an engagement ends, raw training data is deleted from our systems within 90 days unless you have requested it be deleted sooner or transferred to you. Tailored weights are held for as long as you choose to keep them with us.
Deletion requests
Request deletion at any time by writing to support@basisresearch.tech. We complete deletion within 30 days and provide written confirmation.
Your rights
- Access your raw training data and a list of what we hold at any time.
- Correct or supplement the corpus you have provided.
- Request deletion of your raw training data within 30 days.
- Export your tailored weights in a standard format at any time.
- Receive a written DPA that documents every processing activity.
- Choose the region in which your data and weights are stored.
Your data. Your weights. Your model.
Bring your corpus and your standard. We fit, evaluate and deploy a private model, and the weights are yours to keep.