Finetuning Data Policy
This page is the complete data policy for self-serve finetuning. It is designed to answer your compliance questions without a conversation with our team.
Isolation
- Every customer's finetuning data lives under an isolated storage prefix keyed to your account; datasets and artifacts are never comingled across customers.
- Training jobs run in dedicated, per-job directories on isolated compute. Your data is only mounted into containers executing your job.
- Tuned weights are only ever served through your endpoints, authenticated by your API keys.
Where your data physically lives
Two places, both under your account's isolated prefix, both covered by every deletion path below:
| Store | What lives there | Why |
|---|---|---|
| InertialAI object storage | Raw uploads (until auto-deletion), tuned weights, eval + validation reports | The system of record; serves eval exports and local-mode inference |
| Training volume | A per-job working copy of the dataset and the checkpoint the trainer wrote | Mounted into the training container and read by your serving endpoints |
A job delete (or the account-level hard delete) purges the job's directory from both stores; nothing about your data survives in the training infrastructure after deletion.
Retention
| Artifact | Default retention |
|---|---|
| Raw dataset upload | Retained (free), so held-out validation samples stay browsable in the sandbox and re-training needs no re-upload; set retain_uploads: false on the job to delete it when training completes |
| Tuned weights | Retained while the job/endpoint exists |
| Eval report | Retained while the job exists |
| Validation report | Retained while the job exists |
Either way you can hard-delete a job's data — or everything — at any time. A small held-out excerpt stored with the model artifacts powers the sandbox's validation samples and survives an upload opt-out.
Storage is free: there is no charge for retained weights, reports, or opted-in uploads, and no automatic expiry today. If a retention window is ever introduced, it will be announced in advance and existing artifacts will be grandfathered.
Deletion
DELETE /api/v1/finetune/data (or the equivalent dashboard action)
hard-deletes everything: raw uploads, tuned weights, eval artifacts, job
records, and serving endpoints for your account. Serving stops immediately.
Training use
Your datasets are used only to train your finetunes. They are never used to train InertialAI's base models and never shared with other customers.
Region
Training and serving run on secure cloud GPU infrastructure in the general pool by default; region-pinned serving is available in the serving configuration (at the published region rate multiplier).