May 2024 Summaries

Transform v2 is an advanced tool that simplifies the process of detecting and redacting personally identifiable information (PII) in tabular and free text formats using Named Entity Recognition (NER). It offers four robust NER functions - redact_entities, hash_entities, fake_entities, and label_entities. Unlike legacy systems, Transform v2 NER does not restrict users to a predefined set of entities; instead, it allows the listing of any entities needed in the config for detection without requiring prior training. This versatile tool is useful for organizations dealing with unstructured data types like call center transcripts or clinical and doctor notes, where protecting PII is crucial while maintaining maximum utility of the text for downstream applications.

The text discusses generating differentially private synthetic text using Gretel GPT to protect sensitive information in datasets such as customer call logs and medical notes. Differential privacy is a technique that adds calibrated noise during the learning process, reducing the risk of exposing unique linguistic patterns or specific contextual details. The effectiveness of differential privacy fine-tuning is demonstrated using two datasets: augmented-clinical-notes and commonsense-dialogs. Results show that models trained with DP can produce synthetic text attaining comparable Text SQS to those trained without DP, maintaining the quality of the original data while ensuring privacy. Tips for DP fine-tuning are also provided, including suggestions on learning rate, batch size, epochs, dataset size, and compute considerations.

In this blog post, the authors discuss fine-tuning CodeLlama models on Gretel's synthetic Text-to-SQL dataset using AWS SageMaker JumpStart. They demonstrate how to prepare the dataset and create an instruction prompt template for fine-tuning. The fine-tuned model is then evaluated on the BIRD benchmark, showing significant improvements in execution accuracy (EX) and valid efficiency score (VES). This highlights the potential of synthetic datasets in enhancing LLMs for specialized tasks like Text-to-SQL parsing. A SageMaker notebook for this blog post is available, along with Gretel's platform on AWS Marketplace.