A checked training recipe for verifiable reasoning tasks.
What it does
Produces a validated GRPO configuration and reward functions for training behavior with verifiable task success.
When to use it
Use for algorithmically checkable tasks, GRPO reward-function design, or diagnosing a diverging or reward-hacking GRPO run.
How to use it
Give it an RLVR-via-GRPO routing decision and a task verifier; it returns a validated training configuration and reward functions.
What you provide
- Your files
What you need · 4
The recipe and reward-function implementations use Python.
The reference recipe imports GRPOConfig and GRPOTrainer from trl.
The default recipe uses vLLM-backed generation colocated with training.
The default colocated recipe runs generation and training on the same GPU; memory requirements depend on model size.
Access · 0
No additional actions listed.