Company
Post-training for teams who are done collecting for its own sake.
SmallDataRegime.ai exists because the valuable update is usually the small one: a reward written carefully, a few hundred preferences that disagree with the base model, and an eval that can stop the run.
Most post-training stacks assume the scarce resource is compute. In the work we see, the scarce resource is a correct preference. A tired label on an easy example does not make a better policy. A precise correction on a failure does.
We build the system around that constraint. Specs are first-class. Runs have budgets. Releases wait on evals. The hardware can sit in your lab.
Careers
We hire people who have shipped a reward model, a preference interface, or an eval harness, and who would rather delete a bad example than add a thousand more. Write through the contact page with a task you think is stuck in the small-data regime.