Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add redai-infra/Relax --skill nemo-gym-recipe-integrationgit clone --depth 1 https://github.com/redai-infra/RelaxWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/redai-infra/relax/nemo-gym-recipe-integration)<a href="https://agentmods.dev/skills/redai-infra/relax/nemo-gym-recipe-integration"><img src="https://agentmods.dev/badge/skills/redai-infra/relax/nemo-gym-recipe-integration/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/redai-infra/relax/nemo-gym-recipe-integration"><img src="https://agentmods.dev/badge/skills/redai-infra/relax/nemo-gym-recipe-integration.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00073 | $0.02877 |
| Opus 5 | $0.00036 | $0.01438 |
| Sonnet 5 | $0.00015 | $0.00575 |
| Haiku 4.5 | $0.00007 | $0.00288 |
Grade A, and why
nemo-gym-recipe-integration scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 12d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 241 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Integrating a NeMo Gym recipe
Build a new recipe by following the proven Calendar flow and the accumulated PITFAIL records. Keep the user-facing workflow to exactly three steps:
- prepare data;
- start the local NeMo Gym service;
- start remote Relax training.
Non-negotiable rules
- Deliver executable scripts, not a prose-only procedure.
- Run steps 1 and 2 locally. Fix and retry until data, readiness, and the verifier all pass.
- Put only deployment-specific values in
env.sh: image, model path, Gym host/port, shared data path, and optionally remote Ray address. - Write training hyperparameters directly in the training script. Do not create
NEMO_GYM_NUM_ROLLOUT,NEMO_GYM_N_SAMPLES_PER_PROMPT, context-length, batch-size, or parser environment variables. - The training script must contain the complete parameter arrays and final
ray job submit. It may usescripts/run_training.shas the remote job entry andscripts/run_agent_app.shas the thin agent client, but must not delegate to another model training recipe. - Never enable
--no-waitby default. Add it only when the caller explicitly setsRAY_NO_WAIT=1. - Data preparation must create the exact shared file checked by training. A file left only in a Docker volume is not prepared data.
- Training consumes the raw NeMo Gym JSONL. Do not pass an already converted
*_relax.jsonlback throughconvert_dataset.py. - Do not claim success from
/readyz, a Ray JobSUCCEEDEDstate, or a one-sample rollout alone. - A NeMo HTTP 500 must leave a traceback in
docker logs; do not accept access-log-only failures. - Preserve user changes and replace only the named recipe container after verifying its ownership label.
- Never run broad
ray stop,pkill, Docker prune, or sandbox cleanup on a shared system.
Read before editing
Read these files completely because they evolve with real failures:
rg --files examples/nemo_gym_agentic/recipes | rg '/PITFAIL\.md$' | sort
Then read:
- every returned
PITFAIL.md; - the target environment config in the pinned NeMo Gym checkout;
- the target agent, resource server, dataset declaration, verifier, and cleanup behavior;
examples/nemo_gym_agentic/service/Dockerfile;examples/nemo_gym_agentic/scripts/convert_dataset.py;examples/nemo_gym_agentic/scripts/run_training.sh;examples/nemo_gym_agentic/scripts/run_agent_app.sh;- the Calendar recipe as the simple-agent reference;
- Workplace Assistant for stateful tool/resource cleanup;
- R2E-Gym for sandbox, artifact, and multi-process callback propagation.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 12d ago First seen · 241 lines · 73 tokens per session scan A 6164f735da29
nemo-gym-recipe-integration is a skill published in the GitHub repository redai-infra/Relax (580 stars, last pushed 14d ago), licensed Apache-2.0. It adds 73 tokens to every session and 2,877 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
verl-rl-training
Provides guidance for training LLMs with reinforcement learning using verl (Volcano Engine RL). Use when implementing RLHF, GRPO, PPO, or other RL algorithms for LLM post-training at scale with flexible infrastructure backends.
openrlhf-training
High-performance RLHF framework with Ray+vLLM acceleration. Use for PPO, GRPO, RLOO, DPO training of large models (7B-70B+). Built on Ray, vLLM, ZeRO-3. 2× faster than DeepSpeedChat with distributed architecture and GPU resource sharing.
verl-rl-training
Provides guidance for training LLMs with reinforcement learning using verl (Volcano Engine RL). Use when implementing RLHF, GRPO, PPO, or other RL algorithms for LLM post-training at scale with flexible infrastructure backends.
openrlhf-training
High-performance RLHF framework with Ray+vLLM acceleration. Use for PPO, GRPO, RLOO, DPO training of large models (7B-70B+). Built on Ray, vLLM, ZeRO-3. 2× faster than DeepSpeedChat with distributed architecture and GPU resource sharing.
verl-rl-training
Provides guidance for training LLMs with reinforcement learning using verl (Volcano Engine RL). Use when implementing RLHF, GRPO, PPO, or other RL algorithms for LLM post-training at scale with flexible infrastructure backends.
openrlhf-training
High-performance RLHF framework with Ray+vLLM acceleration. Use for PPO, GRPO, RLOO, DPO training of large models (7B-70B+). Built on Ray, vLLM, ZeRO-3. 2× faster than DeepSpeedChat with distributed architecture and GPU resource sharing.