Lexsi hackathon · Track 2b · Circuit-level interventions
A safer Tiny Aya in five languages, audited for side effects
Two LoRA adapters make Tiny Aya Global help with fewer harmful requests. A five-part audit then measures what else changed.
How it was built
- FindCircuitKIT
Score every attention head and MLP for refusal.
- ChangeLoRA
Train only the top 30% of components.
- CompareSafeTune
Steering as the baseline, random components as the control.
- AuditAuditKIT
Ability, safety, five languages, other circuits, hidden errors.
Fewer harmful answers
Lower is better in every chart here. SafeTune steering is the baseline to beat.
Harmful requests answered
Hindi, Arabic, French and Swahili
Harmless prompts refused
Five languages: the cost of refusing more
Harmful requests answered, by language
What the audit caught
The first adapter kept writing after it refused, and its text got less fluent. The second adapter fixes both.
Replies that stop at the refusal
Share of replies to harmful requests. Higher is better.
Perplexity increase on unseen text
Against the base model. Lower is better.
Everything else we measured
Change from the base model in percentage points, with 95% intervals. Left of zero is worse.
Accuracy on reworded facts falls about 4 points. The other checks show no clear change. Four of the five audit parts were run on the adapters; the "other circuits" part was run only on the earlier runs.
Is it the circuit? No.
The same adapter trained on random components does just as well.
Circuit, random or all components: the same result
Harmful requests answered outside English. Each dot is one training run. Lower is better.
The circuit is not small
The top 30% of components by refusal score, the set both adapters were trained on.
Two warnings for anyone testing model edits
Passing the usual check is not enough
Share of runs that pass the usual side-effect check and are still damaged.
Usual check: perplexity within 5% and MMLU within 5 points. Damaged: something we audit moved by 10 points or more. The 10-point line was set after the study.
Counting refusal phrases overstates the language gap
Gap between English and Hindi on the base model, in percentage points.
The data
Mean over training seeds.