What problem does it solve?
Adding a new NVIDIA NeMo Gym environment to the Relax RL framework involves data preparation, local Gym service setup, remote Ray training launch, and verifier validation, with many subtle failure modes. This Skill encodes the proven three-step recipe workflow and accumulated pitfall records so each integration is executable, validated, and reproducible.
Core Features & Use Cases
- Three-Step Recipe Workflow: Generates the seven required recipe files (env.sh, prepare script, Gym launcher, training script, verifier, README, PITFAIL.md) covering data preparation, local Gym service, and direct Ray training submission.
- Validation Ladder: Enforces staged checks from static linting through data integrity, service readiness, network reachability, single-trial verification, and multi-sample training with reward variance and optimizer steps.
- Failure Triage: Maps symptoms like Gateway 404 callbacks, reward-always-zero, and cleanup_unverified to concrete first checks, separating genuine model errors from integration bugs.
- Use Case: When adding a new agentic RL environment (e.g., a stateful tool-use benchmark) to Relax, use this Skill to produce the complete recipe, validate the verifier distinguishes correct and incorrect responses, and confirm a real training run reaches an optimizer step and checkpoint.
Quick Start
Integrate the specified NeMo Gym environment into Relax as a new recipe under examples/nemo_gym_agentic/recipes, following the three-step prepare, serve, and train workflow.
