LLM Evals · Turn quality into a signal
How do you evaluate toxicity and safety?
Short answer
Use policy-defined categories, a mix of classifiers and human review, and measure both harmful output and over-blocking. Safety is a distribution of behaviours, not a single score.
Safety evaluation needs a policy: what is disallowed, what requires caution and what is a legitimate request. Then it needs cases that probe each category, including benign prompts that should not be blocked. A system that refuses everything is safe but useless.
- Cover harmful requests, jailbreaks, prompt injection and sensitive topics.
- Include benign borderline cases to measure false refusals.
- Use classifiers for scale and humans for ambiguous or high-impact cases.
- Report harm categories separately and track refusal rate over time.
Follow-ups you should expect
Interviewers use these to check whether the first answer was understanding or memorisation.
- 1How would you measure over-blocking?
- 2Why is a single toxicity score insufficient?
- 3How often would you red-team the feature?
Say it in the interview
I start from a written policy and build cases for each category, including benign prompts that should not be blocked. Then I combine classifiers for scale with human review for ambiguous cases. I report harm categories separately and track the refusal rate, because over-blocking is a real quality failure. The thresholds themselves are product and legal decisions, so they need a named owner.
Use this as a model, not a script. Saying it in your own words is what makes it survive follow-up pressure.
Related questions
Reading is step one
Now say it out loud.
Understanding an answer and delivering it under pressure are different skills. Rehearse this topic in a live LLM Evals mock interview and get scored feedback on depth, clarity and confidence.