PolicyGrad makes it straightforward to run SFT, DPO and GRPO on open-weight models: define a reward, train a policy, evaluate it, and ship it to inference without stitching together a dozen research repos.
Teams that want a model tuned to their task end up hand-wiring trainers, reward functions, rollout servers and eval harnesses. Runs are expensive, unstable and hard to compare, and the gap between a research recipe and a production model is wide.
Supervised fine-tuning, preference optimization (DPO) and online RL (GRPO) behind one API, with sensible defaults and reproducible configs.
Plug in verifiable rewards, model-graded rubrics or custom Python reward functions, and track every checkpoint against held-out evals.
High-throughput generation built on vLLM for rollouts during training, and export to the same serving stack once a policy is ready.
Start from an open-weight base or instruct model.
Prompts, demonstrations or preference pairs.
Checkers, rubrics or your own scoring code.
Run SFT → DPO → GRPO and compare checkpoints.
Export weights or serve the policy directly.
We're onboarding a small group of early users in 2027. Tell us what you're training and we'll reach out.