Data handling
Baseten doesn’t use your content for model training, fine-tuning, or improvement of any algorithms. What Baseten stores depends on the product:- Model APIs: Requests run under a zero data retention (ZDR) policy. Baseten doesn’t store model inputs or outputs.
- Dedicated Inference: Synchronous requests run under ZDR by default, so Baseten processes your model inputs and outputs without storing them.
- Training and Loops:
- Loops deletes training data and prompts once it processes them.
- Checkpoints from your training jobs and Loops runs stay available to deploy or reuse until you delete them. To delete them, see delete training projects and jobs and delete Loops runs and checkpoints.
- To skip repeated downloads between training jobs, you can turn on the training cache in your job’s configuration. The cache then keeps the files your jobs write to it, and deleting the project permanently deletes them.
- Baseten Delivery Network (BDN): To reduce cold-start time, you can turn on BDN in your model’s configuration. BDN then mirrors your weights to Baseten and caches them near your replicas, so each replica reads from the cache as it starts. Baseten permanently deletes cached weights on request.
- Sandboxes:
- A sandbox keeps its files until it expires or you delete it. Set an expiration policy to control when Baseten deletes the sandbox and its files.
- To keep files after a sandbox is gone, you can mount a volume. The volume then keeps its files until you delete it.
Compliance
Baseten Cloud and single-tenant hosting support SOC 2 Type II, HIPAA, and GDPR compliance programs. See Baseten’s current certifications and attestations in the Baseten Trust Center, and request the SOC 2 Type II report there under a non-disclosure agreement.Access control
- SSO and SCIM: Authenticate users through your identity provider with single sign-on (SSO), and provision accounts and roles with System for Cross-domain Identity Management (SCIM). Available on the Enterprise plan.
- Roles: Assign roles that set who can deploy, manage, and call models.
- Restricted environments: Allow only designated users to change environments such as production.
- API keys: Scope team keys to specific models or environments, and set an expiration when you create a key. Personal keys require an expiration.
- Audit logs: Review configuration changes and lifecycle events for models, Chains, secrets, API keys, environments, and users. On the Enterprise plan, export audit events to destinations such as Amazon S3, Datadog, and Splunk.
- OpenID Connect (OIDC): Give deployments and training jobs short-lived tokens for your cloud resources and Hugging Face repositories in place of long-lived credentials.
- Secrets: Store credentials that Baseten encrypts and injects into your model’s environment at runtime.
Hosting and network controls
- Baseten Cloud: Baseten operates the infrastructure on its multi-cloud fleet.
- Single-tenant: Baseten operates compute and networking dedicated to your organization, with optional private connectivity. Available with custom pricing on the Enterprise plan.
- Self-hosted: The workload plane runs in your own cloud account, and inference payloads stay inside your VPC. Available with custom pricing on the Enterprise plan.
- Regional environments: Constrain replicas to a designated geographic region and route requests through a region-specific endpoint. Baseten configures them on request.
- Egress restrictions: Limit a deployment’s outbound traffic to an allow list of IP ranges and domains. Baseten enables them for your organization on request.