Responsible Use, Safety and Maintenance

You own the safety layer

Hosted providers run safety systems around their models. When you self-host, you are responsible for:

• Input and output filtering appropriate to your users (moderation models or classifiers).
• Preventing misuse: authentication, rate limits, usage policies.
• Protecting against prompt injection in tool-using or RAG features (treat retrieved text as data, least-privilege tools).
• Human review for high-stakes decisions.

Data governance

• Know where prompts and outputs are stored, for how long and who can read them.
• Anonymise or minimise personal data in logs and training sets.
• If you fine-tune on user data, have consent and a legal basis, and be able to retrain if someone asks to be removed.

Licensing and attribution

Track each model's licence and any obligations (attribution, use restrictions, limits on using outputs to train other models). Check again when you upgrade to a new version: licences can change between releases.

Keeping current

Open models improve quickly. Re-run your evaluation set when promising new models appear, and upgrade when the numbers justify it. Pin versions in production and roll out new ones gradually, comparing results.

Supply-chain security

Download models from official sources, verify checksums where provided, prefer safe weight formats (safetensors, GGUF) and keep serving software patched.

Final project

Deploy a private document assistant for a small business: an open-weight model served behind an authenticated Flask API, RAG over their documents with citations, output filtering, usage logging without personal data, an evaluation set with a quality threshold, and a written comparison against a hosted API model including total cost of ownership.

Check yourself

• What safety responsibilities move to you when self-hosting?
• Why re-check licences when upgrading models?
• How do you decide when to upgrade a production model?