Skip to content

Legal

Model Improvement & Data Use

Effective 2026-07-13·Last updated 2026-07-13·HEOSSI (Pte.) Ltd.

This policy explains when interactions may be used to improve Bee. Optional model-improvement processing is off by default. It is separate from processing needed to answer a request, keep an account secure, bill for service, or comply with law.

1. The default

  • Bee does not use Customer Data to pre-train a foundation/base model.
  • Improve Bee is off by default for new and existing accounts. When it is off, conversation exports and identifiable safety-event copies are excluded from adapter, routing, evaluation, and safety-model training pipelines.
  • Enterprise and Enclave Customer Data is excluded unless an authorised customer administrator gives separate written instructions for a tenant-scoped custom-training engagement.

2. What happens when you opt in

  • If you enable Improve Bee in Account > Privacy, selected conversation messages, feedback, and safety-blocked interactions may be copied into a restricted improvement dataset.
  • We use opted-in data to fine-tune Bee adapters, improve request routing and refusal behaviour, build evaluations, diagnose systematic quality failures, and conduct internal Bee Ignite safety research.
  • Opt-in does not authorise us to publish your conversations, train an unrelated third-party model, advertise to you, or expose your content to another customer.

3. Safety and abuse processing

  • Bee may scan requests and outputs to enforce the Acceptable Use Policy whether or not Improve Bee is enabled. That real-time processing is necessary to secure the service and enforce the contract.
  • When Improve Bee is off, the training pipeline does not retain an identifiable research copy of the blocked prompt. We may retain minimised security metadata—time, account id, rule triggered, and request identifiers—for fraud prevention, investigation, and legal compliance under the server-log retention schedule.

4. Data minimisation and protection

  • Improvement datasets use pseudonymous account hashes instead of email addresses where practicable. Research-queue content is access-restricted and encrypted at rest when the platform encryption key is available.
  • Automated and human curation seeks to remove obvious secrets and direct identifiers before a sample is approved for training. Do not submit confidential or sensitive personal data merely because this screening exists.
  • Hugging Face may host restricted improvement datasets as a sub-processor; access is not public. Other training infrastructure is listed in the DPA when it processes personal data.

5. Withdrawal and deletion

  • Turn Improve Bee off at any time. The change applies to future collection without requiring account cancellation and does not affect normal service.
  • To remove prior identifiable improvement copies, email bee-privacy@heossi.com. We will delete or irreversibly de-identify eligible copies within 30 days, subject to legal holds and backup expiry.
  • A trained model cannot ordinarily identify or surgically remove a particular training example. We therefore control the source dataset and future training use; where technically feasible and legally required, we also evaluate model-level remediation.

6. Custom and community training

  • Tenant-scoped custom training requires documented customer instructions and remains governed by the DPA. It does not place the customer's data in Bee's shared improvement dataset.
  • Publishing an adapter to a community surface is a separate, deliberate disclosure. The publisher must own or have permission for all training material and must not publish personal, confidential, infringing, or unlawful content.

Questions about this document? Contact bee-privacy@heossi.com. Service of process: bee-legal@heossi.com (HEOSSI (Pte.) Ltd., Singapore).

Counter-signed copies on request. The text on this page is the canonical published version. For procurement teams that need a counter-signed copy of the Terms, DPA, or Order Form, email bee-legal@heossi.com. Where there is conflict between this page and an executed counter-signed agreement, the counter-signed agreement controls.

Terms of ServicePrivacy PolicyData Processing AddendumAcceptable Use PolicySecurity PracticesBEE for Startups Program TermsPartner Program TermsVC Partner Program TermsCookie PolicyAI Transparency & DisclosureAccessibility StatementCopyright & DMCA PolicyGovernment & Law-Enforcement Data RequestsRegional Terms & Mandatory RightsContent Moderation, Notices & AppealsDeveloper & API TermsModel Governance, Training Sources & Copyright
HEOSSI (Pte.) Ltd. · Singapore · heossi.comSee evidence index →