purvansh01/tiny-aya-global-safetypulse-repaired
Text Generation • 3B • Updated • 336 • 1
Repaired model and Tiny Aya Global LoRA adapters from SafetyPulse (Lexsi hackathon, Track 2b): safety drift in fine-tuning, monitored by CIS.
Note The repaired model: drifted Alpaca fine-tune + SafeTune safety-vector restore. ASR 24.4% -> 11.8%, GSM8K 56.7% -> 71.3%; CIS FAIL -> PASS.
Note The repaired model in 4-bit NF4 (2.5 GB), with a paired audit: part of the repair is lost (ASR 10.9% → 16.8%).
Note Alpaca fine-tunes (4 seeds x 11 checkpoints): benign tuning that weakens refusals; warning card.
Note Same run after CuratorKIT curation: drift unchanged, CIS still warns early.
Note GSM8K domain fine-tunes (3 seeds): safety kept, monitor WARN only.
Note Negative control (3 seeds): safety-preserving data, monitor PASS.