Datology

DatologyAI is a frontier data research lab that builds a data refinery for AI model builders. Its Curation Studio takes an organization's own data and turns it into higher-quality training data for mid-training open models or pre-training foundation models, deployed on the customer's private cloud. The company says Curation Studio is now available to every AI team.

What It Is

Datology is a data curation platform for teams training their own models. It does not sell data or source tokens; it improves the data a team already has, whether proprietary enterprise data, open public web data, or licensed data from partners and providers. The About page describes Curation Studio as a guided experience with presets and automation on top of Datology's curation algorithms, aimed at AI teams from beginners to experts.

The Four-Stage Pipeline

The refinery works in four stages:

• Clean: removes poorly formatted, empty, short, and eval-contaminated data.
• Calibrate: tailors data by quality, task, and taxonomy to the use case.
• Create: generates synthetic data to maximize task relevance and diversity.
• Compose: mixes data sources into staged training datasets.

The output is a curated training dataset for building a better model.

Compute Multiplier and Case Studies

Datology states that better data quality raises the signal per token in training, which it calls a compute multiplier, and the homepage banner claims a proven 6x compute multiplier. It publishes two case studies. In the first, Thomson Reuters mid-trained against a proprietary legal corpus, which Datology reports gave +5% on Legal Bench and +2.5% on general evals. In the second, Arcee built a 398B-parameter open-weights MoE model with 13B active parameters, using 17T tokens curated by Datology.

Background

The company is led by co-founders Ari Morcos (CEO), Bogdan Gaza (CTO), and Matthew Leavitt. It also publishes research updates, such as DataSmith for automating data research.