Toloka AI BV, a Dutch provider of datasets for artificial intelligence projects, today announced that it closed a $72 million investment.

The round was led by Bezos Expeditions, a fund affiliated with former Amazon.com Inc. Chief Executive Officer Jeff Bezos. Shopify Inc. Chief Technology Officer Mikhail Parakhin participated as well.

Toloka is a unit of Nasdaq-listed AI infrastructure provider Nebius Group NV. The latter company, which is based in the Netherlands, was originally the parent organization of Russian search engine Yandex. Nebius was suspended from the Nasdaq in 2023, offloaded its Russian holdings the following year and subsequently resumed trading on the stock exchange.

Today, the company’s main focus is operating a public cloud optimized for AI workloads. Nebius raised $700 million from Nvidia Corp., Accel and other backers in December to enhance the infrastructure that powers the platform. It plans to build a new data center in New Jersey and expand several existing facilities.

Following today’s investment, Nebius will no longer have majority voting control of Toloka and won’t include the latter company’s earnings in its quarterly results. It will retain a “significant majority economic stake.”

Amsterdam-based Toloka provides custom datasets that AI companies can use to train their models. It relies on a network of more than 200,000 annotators and other professionals to create those datasets. According to Toloka, the experts in its network can create AI training materials in more than 40 languages.

After training a neural network, companies often apply optimizations to better align its output with user preferences. A retailer, for example, may prefer that each AI-generated shopping recommendation mention at least two products from its catalog. The process of aligning an AI model with user requirements involves supplying it with additional training data.

Toloka says that its platform makes it easier for developers to source the necessary training data. It supports RLHF and DPO, the two most widely-used methods of aligning AI model output with user preferences. In RLHF projects, developers fine-tune an AI model’s output using a second neural network trained to understand user preferences. DPO, in contrast, doesn’t rely on a second neural network, which often makes the method more cost-efficient.

Training AI programming assistants is another task that Toloka promises to ease. The company can provide code samples in more than a dozen programming languages. Users can customize details such as the number of code files in a training dataset, the tasks those files perform and the development frameworks they use.

Before releasing an AI model to production, developers test that it’s safe using a method known as red-teaming. They supply the model with a large number of prompts that mimic malicious input and evaluate how it responds. According to Toloka, AI teams can source such prompts through its platform to speed up red-teaming initiatives.

The company’s platform is used by Anthropic PBC, well-funded coding assistant startup Poolside and other major players in the AI market. Toloka stated that today’s investment will facilitate the “significant and rapid scaling” of its business. 

Image: Toloka