Training data for frontier models.

Variable Ratio is a research lab. We build reinforcement learning environments, pre-training corpora, and supervised fine-tuning data for teams training their own models.

Get in touch →

RL environments

Task suites with verifiable rewards, built to spec. Agentic tool use, software engineering, research workflows, long-horizon planning.

Pre-training data

Sourced, filtered, and deduplicated corpora at scale, with provenance and licensing documented for every shard.

SFT data

Demonstrations and preference data written by domain experts, with review passes and rubrics you can inspect.