November 2024 Summaries

The text discusses how large language models (LLMs) can be leveraged to transform natural language into functional Python code for FinTech applications, thereby lowering technical barriers and accelerating innovation in the industry. It explains the process of creating a synthetic Text-to-Python dataset for FinTech using Gretel Navigator SDK's Data Designer mode. The dataset is carefully crafted with domain-specific terminology and scenarios to enable LLMs to generate precise, actionable code. The text also provides details on how to set up the workflow, build the dataset configuration, run the pipeline, and validate data quality. Finally, it introduces Gretel's Synthetic Text-to-Python Dataset for FinTech, a publicly available collection of 25,000 records tailored to support various FinTech coding applications.

Gretel's PII Replay is a new privacy metric that identifies instances of sensitive information found in original training data and counts how often those values appear in synthetic output. This tool works alongside Membership Inference Protection and Attribute Inference Protection, ensuring your synthetic data remains private by design. By leveraging Gretel Transform to identify and classify instances of PII in the original training data, users can now easily see whether any of the original PII is showing up in their synthetic data. Strategies to minimize PII Replay include using Transform before generating synthetics, choosing a model designed to minimize PII replay, leveraging differential privacy, pre-processing to remove unnecessary columns, and using pre- and post-processing strategies strategically.

Navigator Fine Tuning (NavFT) is a model used by Gretel for generating synthetic tabular datasets containing various types of fields. The company has now enabled fine-tuning with differential privacy (DP), which provides strong, formal guarantees against the leakage of sensitive information about the original dataset. Gretel supports three models with DP: NavFT, Gretel GPT, and Tabular DP. NavFT is ideal for tabular datasets with mixed column types, while Gretel GPT is best for free-text-only datasets, and Tabular DP is suited for numerical and categorical-only datasets. The use of differential privacy can enhance resilience to attacks but may reduce synthetic data quality, requiring a balance between privacy and utility.

Navigator Fine Tuning (NavFT) is a model used by Gretel for generating synthetic tabular datasets containing various types of fields. The company has now enabled fine-tuning with differential privacy (DP), which provides strong, formal guarantees against the leakage of sensitive information about the original dataset. Gretel supports three models with DP: NavFT, Gretel GPT, and Tabular DP. NavFT is ideal for tabular datasets with mixed column types, while Gretel GPT is best for free-text-only datasets, and Tabular DP is suited for numerical and categorical-only datasets. The use of differential privacy often reduces synthetic data quality, requiring a balance between privacy and utility. Experiments show that employing DP boosts the Data Privacy Score while decreasing the Synthetic Quality Score.

The Navigator Data Designer's sample-to-dataset capability allows users to generate thousands of records from just a few samples by automatically expanding and exploring the problem domain as a human expert would. This feature maintains the exact format and context of the data, including style, template, and code, while also providing automatic data cleanup and standardization out of the box. The sample-to-dataset workflow is designed for creating diverse and rich datasets tailored to specific AI training/fine-tuning needs, addressing class imbalances, and increasing complexity and richness of data.

Financial institutions face challenges in utilizing valuable data due to regulatory constraints and privacy concerns. A reference architecture addressing these issues includes a Data Governance Layer for maintaining control over sensitive data, a Privacy-Preserving Synthetic Data Generation Layer powered by Gretel's synthetic data platform, and an AI Development Layer built on Azure OpenAI Services. This approach enables secure data sharing across teams while ensuring model safety through privacy attack simulations and robust data quality reporting. By combining differential privacy capabilities with Azure OpenAI Services' language models, organizations can generate highly accurate synthetic datasets that protect sensitive financial information and prevent PII exposure.

Gretel Navigator's new Data Designer tool simplifies synthetic data creation for AI development teams, enabling them to generate high-quality training datasets without compromising sensitive information. With features like speed, quality, simplicity, scale, and data-centric AI, the tool helps users build diverse datasets tailored to specific business needs. It also offers a range of blueprints for solving specific use cases, such as Text-to-Python and Text-to-SQL datasets. The Data Designer is currently in early preview, with opportunities for users to participate and contribute to its development.