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 nebius/nebius-physical-ai --skill cosmos3-post-traininggit clone --depth 1 https://github.com/nebius/nebius-physical-aiWrote 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/nebius/nebius-physical-ai/cosmos3-post-training)<a href="https://agentmods.dev/skills/nebius/nebius-physical-ai/cosmos3-post-training"><img src="https://agentmods.dev/badge/skills/nebius/nebius-physical-ai/cosmos3-post-training/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/nebius/nebius-physical-ai/cosmos3-post-training"><img src="https://agentmods.dev/badge/skills/nebius/nebius-physical-ai/cosmos3-post-training.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 | $0.00051 | $0.00889 |
| Opus 5 | $0.00026 | $0.00445 |
| Sonnet 5 | $0.00010 | $0.00178 |
| Haiku 4.5 | $0.00005 | $0.00089 |
Grade A, and why
cosmos3-post-training 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 4d 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 — 89 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Cosmos3 Post-Training
Source And Attribution
Adapted from NVIDIA cosmos-framework
skills/workflows/cosmos3-post-training/SKILL.md.
Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. Used under OpenMDW-1.1.
See skills/LICENSE-NVIDIA-COSMOS3-OPENMDW-1.1 and
skills/NOTICE-NVIDIA-COSMOS3.
When To Use
Use this skill when the user asks how Cosmos3 SFT works, how to review future post-training support, where upstream recipes live, how to validate training configs, or whether an NPA change should expose post-training.
For current NPA, treat Cosmos3 post-training as guidance and planning unless a real executable workflow is implemented and tested. Do not add a Cosmos skill-display subcommand or a SkyPilot YAML whose only purpose is to make this agent skill runnable.
Current NPA Boundary
Retained real Cosmos3 workflows:
workflows/testing/cosmos-fetch.yamlworkflows/testing/cosmos3-text-to-image.yaml
Current NPA Cosmos commands such as npa workbench cosmos train cover the
existing Cosmos workbench/serverless training surface, not a proven Cosmos3 SFT
workflow. Do not present that as Cosmos3 post-training unless implementation and
tests explicitly support it.
Upstream Post-Training Map
In a clone of https://github.com/NVIDIA/cosmos-framework.git, inspect:
| Need | Upstream path |
|---|---|
| Training guide | docs/training.md |
| Dataset JSONL/captioning guide | docs/dataset_jsonl.md |
| SFT recipes | examples/toml/sft_config/<recipe>.toml |
| Paired recipe launchers | examples/launch_sft_<recipe>.sh |
| Common launcher helper | examples/_sft_launcher_common.sh |
| Training script | cosmos_framework/scripts/train.py |
| DCP conversion | cosmos_framework/scripts/convert_model_to_dcp.py |
| HF export | cosmos_framework/scripts/export_model.py |
| TOML schema | cosmos_framework/configs/toml_config/sft_config.py |
Planning Checklist
When reviewing or designing NPA Cosmos3 post-training support:
- Define the exact executable outcome: config validation, dry run, training, checkpoint conversion, export, or inference from a trained checkpoint.
- Require explicit dataset, base checkpoint, and Wan VAE paths where the upstream recipe requires them.
- Keep training extras explicit:
cu130-trainorcu128-train. - Validate TOML/schema behavior with upstream
train.py --dryrunbefore claiming training support. - Use temporary or user-selected output roots, not repository paths.
- Preserve redaction for Hugging Face, GitHub, NGC, S3, and any other secret env values.
- Add tests that prove NPA maps inputs into a real executable workflow. Do not use tests that only prove an agent skill can be listed or displayed by a CLI.
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.
- 4d ago Changed 57aa5df1f5a0
- 9d ago First seen · 89 lines · 51 tokens per session scan A 9c773032d37e
cosmos3-post-training is a skill published in the GitHub repository nebius/nebius-physical-ai (29 stars, last pushed today), licensed Apache-2.0. It adds 51 tokens to every session and 889 once invoked, about $0.0003 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 2026-09-03.
Other skills, from other repositories
llm-app-patterns
Production-ready patterns for building LLM applications. Covers RAG pipelines, agent architectures, prompt IDEs, and LLMOps monitoring. Use when designing AI applications, implementing RAG, building agents, or setting up LLM observability.
prompt-optimization
Improve a prompt on the evaluations workbench through a measured loop. Score the baseline first, then duplicate the target column, form a hypothesis from failing rows, edit the copy's prompt draft, run, compare pass rate and cost, and repeat until the numbers hold. Use when the user asks to optimize or improve a…
enhance-prompt
Transforms vague UI ideas into polished, Stitch-optimized prompts. Enhances specificity, adds UI/UX keywords, injects design system context, and structures output for better generation results.
prompt-engineer
Writes, refactors, and evaluates prompts for LLMs — generating optimized prompt templates, structured output schemas, evaluation rubrics, and test suites. Use when designing prompts for new LLM applications, refactoring existing prompts for better accuracy or token efficiency, implementing chain-of-thought or few-shot…
seedance-vocab-en
This skill should be used when an English Seedance 2.0 prompt needs clearer production wording, less generic prose, or precise vocabulary for camera, lighting, motion, VFX, audio, and constraints. Route blocked prompts through seedance-filter for context and boundary review.
ideogram4
Prompting patterns for Ideogram 4 text-to-image — best-in-class in-image text rendering and exact color/layout control via structured JSON captions. Use when generating images that need legible on-image text (title cards, thumbnails, logos, signage, CTAs), precise brand colors, or controlled spatial layout. Triggers…