Safety fine-tuning
Also called Safety Fine-Tuning
The fine-tuning stage that teaches a model to avoid harmful, unsafe, or policy-violating outputs.
Think of it like
A driving instructor's final lessons — not new driving skill, but the judgment to not do dangerous things.
Example
After general instruction tuning, a model gets a safety pass using refusal examples and red-team data to close known failure modes.
How it actually works
Safety fine-tuning layers guardrails onto an already-capable model using curated examples, preference data, and adversarial red-team cases. It teaches refusals, careful handling of sensitive topics, and resistance to manipulation. It is never perfect — jailbreaks keep finding gaps — and pushing it hard incurs the alignment tax, so it is a continual balancing act rather than a one-time fix.
For product teams
The stage that makes a model shippable to the public — and the source of most "why won't it" complaints.
For engineers
Post-SFT tuning on safety examples, preferences, and red-team data to reduce harmful outputs.
Related
- Refusal training — A specific skill it instills.
- Red-teaming data — The adversarial data it uses.
- Alignment tax — The cost it can impose.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome