How does Cohere source training data?

Cohere's EU AI Act training summary#

Cohere's public training content summary for Command A+ and related models describes text data as sourced from publicly available content and proprietary datasets accessed through partnerships or developed by Cohere, covering content such as general web content, technical documentation, code, education, government, science and technology material.

The same summary describes image training data as sourced from publicly available content and proprietary datasets, with a focus on enterprise-relevant content such as tables, charts, PDFs and infographics.

Cohere's web crawler policy#

Cohere's documentation states that, as of its publication, Cohere does not use Cohere-branded bots or user agents to crawl or scrape web content for training generative AI foundation models, and that it would identify any such bots if it began doing so in the future.

Transparency index findings#

Cohere publishes multilingual coverage figures and data collection cutoff dates in its EU AI Act summaries, for example stating that one Command A+ model family's training data was collected up to around mid-2026 and spans at least 48 languages.

What this means for a business#

A business considering API access to Cohere's enterprise-focused Command models should review Cohere's public training content summaries directly, since Cohere discloses more granular modality and language detail there than in general marketing material.

SourceX and Cohere#

SourceX is not affiliated with Cohere. This page summarizes public reports; it does not mean Cohere buys data through SourceX.