Skip to main content
pre.dev offers AI labs two kinds of training data. Both come from private production codebases that pre.dev licenses directly from the teams that built them.

RL Tasks

Hard, verified coding tasks for post-training, built from private production code no model has seen. One real commit becomes one task: a Harbor task container at the commit’s parent, an instruction, a graded verifier, and a reference solution. Every task clears QC checks for solvability and for gaming before solver runs place it in a difficulty band. You receive tasks, not repository access. See RL Tasks for the bundle layout, task.toml, the reward, and the QC checks.

Codebases

Licensed private codebases with full git history, cleaned and delivered for training. Each repository ships with every branch and tag, its pull requests with reviews, and a README that explains how to rebuild each pull request as a task. Secrets and personal data are removed from every version of every file, and contributors are pseudonymized. See Codebases for the zip layout, the pull request record, the catalog sheet, and pricing.

How buying works

  1. Tell us what you’re training. The short form on pre.dev/labs or pre.dev/rl-tasks asks for your name, work email, company, and what you’re training.
  2. Book a 30-minute call with the founders. Sample tasks, verifiers, QC evidence, and model results are walked through with your team on the call, and a set is scoped to your target domains and languages.
  3. Start with a paid pilot. Pricing is set per engagement. Licenses are non-exclusive by default.
Trust and data handling covers the privacy checks, what is removed before delivery, and how licensing works.