Datology

Datology turns an organization's data into higher-quality training data for mid-training open models or pre-training its own foundation models. The pipeline runs on your data inside your private cloud. It is aimed at teams that build models and want more from the compute they already have.

The refinery works in four stages. Clean removes poorly formatted, empty, short and eval-contaminated data, Calibrate tailors data by quality, task and taxonomy, Create generates synthetic data for relevance and diversity, and Compose mixes sources into staged training sets.

Inputs can be proprietary enterprise data, public web data or licensed data from partners. Datology says it does not sell data or source tokens. Its case studies describe mid-training on a legal corpus that lifted benchmark scores while using a small share of base pre-training compute.