June 2023 Summaries

Gretel has introduced a Synthetic Text Data Quality Report that measures semantic and structural similarity between AI-generated text and training text in 50 languages. The report includes the Text SQS, which estimates how well the generated synthetic data maintains the same semantic and structural properties as the original dataset. This score can be viewed as a utility or confidence score for drawing scientific conclusions from the synthetic dataset. The report compares Amazon product reviews with synthetic text from Gretel's GPT-x model. It provides recommendations based on the Text SQS, helping users understand its implications and how to improve it if necessary.

Gretel has launched its privacy-enhancing tools and generative AI models on Google Cloud Marketplace, making it easier for enterprises in the Google Cloud ecosystem to generate high-fidelity synthetic versions of their sensitive datasets. These synthetic datasets are statistically accurate, infinitely scalable, and privacy-protected. Gretel's platform features multimodal synthetic data generation, tunable privacy & utility, and quality metrics & analysis reporting. By offering API-driven generative AI modeling capabilities that focus on core use cases for MLOps and LLMOps teams, Gretel enables teams to quickly experiment and innovate safely.

In this blog post, we explore fine-tuning a multi-billion parameter large language model (LLM) using Gretel GPT to answer questions as if it were any of your favorite Twitter personalities. We walk through the steps of training the model with recent tweets from popular Twitter users and provide examples of how the adapted model can be used to generate text reflecting each personality's tone and style. The project is inspired by Boris Dayma's HuggingTweets, which uses a similar approach for generating synthetic tweets. We demonstrate how to compile a list of Twitter users, preprocess their tweets, create a training set, configure the Gretel training process, initiate model training, and generate queries using the trained model. Finally, we examine the results and discuss potential improvements and future directions for this experiment.

Gretel's high-quality synthetic data generation tools are now available in the AWS Marketplace, allowing customers to safely train custom AI models on proprietary data while meeting privacy compliance requirements and unlocking secure data access and collaboration opportunities. The demand for safe, accurate, and timely data has increased with the integration of large language models (LLMs) into services and scaling of generative AI applications. Gretel's comprehensive solution combines privacy-enhancing technologies with cutting-edge generative models to enable anyone to create high-quality, domain-specific synthetic data on demand. Synthetic data is helping teams mitigate privacy and security risks, augment limited data supplies, simulate edge cases, and comply with various privacy regulations.

Gretel, a multi-modal synthetic data platform, has received its Service Organization Control (SOC) 2 Type 1 certification, confirming that its security safeguards can support customers with strict security requirements. The company places significant emphasis on privacy and security, as its founders have spent much of their careers building resilient information ecosystems against cybersecurity and privacy attacks. SOC 2 is a set of criteria developed by the American Institute of Certified Public Accountants (AICPA) that defines security standards for service organizations, covering best practices in security, availability, processing integrity, and confidentiality. The certification helps build trust and confidence among customers, employees, partners, and the public, which is critical for modern businesses developing data-driven products and services.

Gretel GPT, an API for generating synthetic natural language text using Large Language Models (LLMs), is set to receive a series of updates over the next few months. The platform enables fine-tuning state-of-the-art LLMs on enterprise data and supports models like MosaicML's mpt-7b, TogetherComputer's RedPajama 3B model, and StabilityAI’s stable LM 3B model. Gretel GPT also offers PEFT (Parameter-Efficient Fine-Tuning) and LORA (Low-Rank Adaptation of Large Language Models) for supported models to create adapted versions with reduced training times and costs. The platform's auto-parameters feature automatically recommends the best configurations for a dataset, based on training hundreds of thousands of models via Gretel’s cloud service.