October 2023 Summaries

Gretel has streamlined its Python SDK by introducing a high-level interface designed to be simple, flexible, and intuitive for users. With the new interface, users can save time and write 3x less code than before. The updated SDK includes features such as more thoughtful logging, helper methods for assessing synthetic data quality, and a streamlined way to dynamically customize model configurations. Users can now train state-of-the-art deep generative models in just a few lines of code and generate on-demand synthetic data with ease. Gretel SDK Blueprints are available as a resource for users to learn the basics and advanced features of the new interface.

In this blog, the author demonstrates how to prompt Meta's Llama-2 chat model with a series of 250 questions from GSM8k dataset using Gretel's platform. Batch inference is introduced as an efficient method for handling large sets of prompts, allowing organizations to leverage LLMs more effectively across extensive datasets. The tutorial includes setting up the development environment, loading and preparing the dataset, initializing the Llama-2 model, and batch prompting the model. Gretel's platform handles scaling out requests, monitoring progress, assessing synthetic text quality, and fine-tuning models like Llama 2 on custom data. The full code for this tutorial is available on Colab and GitHub.