Add Data Leakage Assessment

[dt-ahmed-touila]

🚀 Feature Request

Measure data leakage based on a subset of samples/conversations from the training/fine-tuning dataset

🔈 Motivation

To efficiently apply generative models to specific domain/application (legal field is a good example) companies turn to fine-tuning / alignment on proprietary/confidential dataset. Once the fine-tuning is done and given the nature of the models, there is a significant risk of training data leakage on inference. This leakage is due to two capacities in LLMs; Memorization and Association.

🛰 Alternatives

• Adhoc implementations 🐰

[luca-martial]

Hi @dt-ahmed-touila, thanks for sharing this request! Since we're blackbox application-focused we don't currently develop tooling for training/fine-tuning related evaluation. We'll re-open this if we get to that domain one day.