Skip to content
← Projects

SafetyPulse

A safety regression gate for LLM fine-tuning

Domain
AI Safety
Status
Active
Year
2026
Stars
★ 5
Built for
Lexsi Labs AI Safety & Alignment Hackathon · Track 2b

SafetyPulse checks whether a model's internal refusal mechanism is still intact during fine-tuning, using a Circuit Integrity Score (CIS) computed from 50 prompt pairs. It gates each checkpoint PASS/WARN/FAIL, repairs failing models with SafeTune, and verifies the repair with a paired audit (AuditKIT).

The problem

Fine-tuning an aligned model on harmless data can quietly weaken its refusals, and full red-team audits of every checkpoint are expensive, so teams skip them.

How it works

  1. 01

    Compute a Circuit Integrity Score from 50 prompt pairs at every checkpoint

  2. 02

    Decide PASS / WARN / FAIL with thresholds fixed in advance and validated on held-out runs

  3. 03

    Repair failing models via SafeTune safety-vector restore

  4. 04

    Verify repairs with a paired McNemar audit (AuditKIT); CuratorKIT and CircuitKIT for data curation and cross-checks

Results

As reported in the repository

14.3% → 25.2% on 119 HarmBench test promptsASR rise from benign Alpaca tuning
Spearman ρ = −0.80 [−0.85, −0.71]CIS vs ASR correlation
25.2% → 12.3% (base 14.3%)SafeTune repair (ASR)
55.1% → 70.7% (base 71.3%)SafeTune repair (GSM8K)
0.95AUROC (cis_sep, post hoc)
55 GPU-s per checkpoint (≈ $0.008)CIS check cost (T4)

Watch

SafetyPulse — demo

Figures

SafetyPulse figureSafetyPulse figureSafetyPulse figure

Same domain