Amusing Abliteration

Abliteration on LLMs is the act of removing guardrails - here I show how to make Llama 3.1 'less kind and good' with questions around explosives, financial restructuring advice, rude jokes and security vulnerabilities. I'm interested in the question - whilst guardrails stop us asking 'awkward questions', what other answers are watered down such that we don't get useful responses?
Created as an outcome of my playgroup research days: https://www.linkedin.com/feed/update/urn:li:activity:7396293087674933248/

Transcript

1. Amusing Abliterations PyDataLondon 2025-12 lightning talk @IanOzsvald – ianozsvald.com
2. At playgroup we talked about humour generation I wondered if ‘abliteration’ – removing safeguards, was a good idea It was The “why” By [ian]@ianozsvald[.com] Ian Ozsvald
3. By [ian]@ianozsvald[.com] Ian Ozsvald Guardrails prevent naughty stuff
4. By [ian]@ianozsvald[.com] Ian Ozsvald Abliteration removes guardrails <- This is the same underlying model, no extra information added
5. By [ian]@ianozsvald[.com] Ian Ozsvald System exploits too
6. By [ian]@ianozsvald[.com] Ian Ozsvald Coarse humour? These safe jokes appear in Google e.g. in reddit/r/DadJokes
7. By [ian]@ianozsvald[.com] Ian Ozsvald I can't tell you what it said! !!CENSORED!! Coarse humour! :-( Unlike dad jokes I made at playgroup, this joke didn't appear in google searches
8. By [ian]@ianozsvald[.com] Ian Ozsvald Private equity – which is abliterated?
9. What is ‘abliteration’? LMStudio (/ollamma etc) What answers do you miss due to guardrails? Next steps: By [ian]@ianozsvald[.com] Ian Ozsvald