Research · Advanced
Evaluating safety and alignment transfer across distilled models
Run matched-pair evaluation benchmarks across teacher and student models to audit alignment shifts.
- When something happensNew model checkpoint logged
- ThenExecute matched-pair evaluation suite
- ResultPresent divergence summary for review
Audit whether fine-tuning or distillation transfers unintended safety filters or behavioral changes. The bot runs matched prompt sets through baseline and distilled models, coordinates automated evaluation judges, and aggregates divergence scores into structured reports.
Prompt to start with
BotBentoRecommended
When evaluation prompts are updated or a model run completes, execute the benchmark suite in the terminal against both baseline and student checkpoints. Compute score distributions across matched evaluation pairs and compile divergence metrics. Format summary tables into documentation. Always request user approval before publishing reports, opening pull requests on the repository, or pushing result logs.