Case Study: Gretel achieves 10x experimentation velocity with Weights & Biases

Gretel achieves 10x experimentation velocity with Weights & Biases

Gretel, a leading synthetic data platform, faced challenges in efficiently fine-tuning the large language models (SLMs) and LLMs used in its Navigator AI system. The process of curating and training on massive, web-crawled datasets comprising trillions of tokens was slow, difficult, and prohibitively expensive, hampering their experimentation velocity. To overcome this bottleneck, Gretel turned to the experiment tracking and evaluation tools from Weights & Biases.

By implementing Weights & Biases, Gretel was able to leverage synthetic data to reduce its training dataset size by 1000x, from a trillion to a billion tokens, while maintaining model quality. This efficiency gain, combined with W&B's logging capabilities, allowed the team to rapidly identify gaps, make adjustments, and launch new experiment batches every few days. The solution dramatically accelerated their workflow, increasing from 3-5 experiments per week to completing 250 experiments in just 45 days, achieving a 10x increase in experimentation velocity.

Gretel