Exabite

Overview

Exabite, formerly Calaveras, supplies pre-training and post-training data for AI developers. Its public catalog overview spans web text, documents, video, supervised training, and reinforcement-learning-related data. The company describes both existing datasets and the ability to source additional material for customer needs.

Its offering has three main components: an off-the-shelf catalog, web-data collection infrastructure, and work on novel or hard-to-find datasets. Exabite emphasizes scale and material beyond commonly used sources. Claims about corpus size, uniqueness, customer reach, and relative cost are company-reported and should be assessed against the actual dataset being offered.

The former Calaveras identity is useful when following older references, but the relevant purchasing unit is still a specific collection with documented contents and terms. A general claim of large-scale supply does not establish that every record is useful, new to the buyer, or appropriate for a particular model.

Where it may be useful

Our assessment is that broad data sourcing is most valuable when a training team can identify a coverage gap or a measurable need for additional diversity. For example, a document model may need unusual layouts and mixed text-image pages, while a language model may need material from underrepresented domains. These are different requirements even if both are described as more data.

Novelty should also be measured relative to the buyer's existing corpus. A collection can be absent from one public source while still overlap heavily with material already acquired elsewhere. Deduplication and sampling are therefore part of assessing value.

Evaluating fit

Request samples, source distributions, collection dates, and documentation of transformations. Compare usable records after filtering rather than headline storage volume. For multimodal material, inspect alignment and whether essential metadata survives delivery. Ask how overlap, corrupted files, and low-quality records affect acceptance and pricing. Establish the provenance and permitted uses of the offered material without treating a marketing description of collection practices as a substitute for dataset documentation. A pilot should test the collection against a defined training objective, with independent evaluation showing whether the added data improves relevant capabilities.