Home/Case Studies/Consumer lender proves that fine-tuning did not strip safety behaviour or leak customer data
Consumer lender proves that fine-tuning did not strip safety behaviour or leak customer data
The lender fine-tuned an open-weight model on customer interactions and built a validation gate that checks for safety regression and training data memorisation before any release.
0memorised customer records extractable in 50,000 probes
98%refusal robustness retained after fine-tuning versus the base model
2open-weight releases rejected on tamper-resistance review
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.
