Canonical: https://botbento.com/usecases/evaluating-safety-and-alignment-transfer-across-distilled/
Format: Markdown representation of the public HTML page.

[Home](/) / [Use cases](/usecases/) / [Research](/usecases/?category=research)

Research · Advanced

# Evaluating safety and alignment transfer across distilled models

Run matched-pair evaluation benchmarks across teacher and student models to audit alignment shifts.

[Use with BotBento ](https://app.botbento.com/?ref=usecases&utm_source=botbento.com&utm_medium=usecases&utm_campaign=evaluating-safety-and-alignment-transfer-across-distilled)Copy promptShare[Markdown](/text/usecases/evaluating-safety-and-alignment-transfer-across-distilled/index.md)

Opens BotBento. Paste the prompt into a bot; nothing is preconfigured.

- 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.
```

Copy

## Tools it uses

TerminalDocsGitHub

## Works with

BotBento

## Tags

- ai
- evaluation
- benchmarking
- alignment
- distillation

## Source

[Show HN: Distilling DeepSeek into GPT-OSS doesn't transfer censorship. Try it](https://news.ycombinator.com/item?id=49113599)
Hacker News · checked 2026-10-15

Summarized in our own words. Any results described there are the authors’ own, not measured by BotBento.

By BotBento Editorial · Added 2026-10-15 · Updated 2026-10-15 · [Suggest a correction](/contact/)

## Related use cases

-

Researchintermediate

## [Benchmarking OCR and vision models on documents](/usecases/benchmarking-ocr-and-vision-models-on-documents/)

Compare text extraction accuracy and formatting across multiple vision models side by side

PDF filesSpreadsheetTerminal
-

Researchintermediate

## [Auditing entity recognition and hallucinations across models](/usecases/auditing-entity-recognition-and-hallucinations-across-models/)

Query public models to evaluate how accurately your identity or brand is represented

BrowserDocsChat
-

Researchadvanced

## [Automated layer modification and benchmark evaluator](/usecases/automated-layer-modification-and-benchmark-evaluator/)

Run systematic layer routing experiments and benchmark model performance across test suites

TerminalSpreadsheetNotes