Legal asks you to remove one customer’s records from an...

The core technical challenge is that model weights are not a database; they do not support row-level deletion. Attempting to ‘unlearn’ specific customer records requires knowing exactly which sequences were memorized, which is not feasible. The only reliable options are to retrain the model from a cleaned dataset or to shift the architecture to RAG, where data can be deleted from the datastore. Note that even after deleting the datastore row, one must consider if the base model memorized the information during pre-training.