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 NVIDIA-TAO/tao-skill-bank --skill tao-train-single-stepgit clone --depth 1 https://github.com/NVIDIA-TAO/tao-skill-bankWrote 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/nvidia-tao/tao-skill-bank/tao-train-single-step)<a href="https://agentmods.dev/skills/nvidia-tao/tao-skill-bank/tao-train-single-step"><img src="https://agentmods.dev/badge/skills/nvidia-tao/tao-skill-bank/tao-train-single-step/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/nvidia-tao/tao-skill-bank/tao-train-single-step"><img src="https://agentmods.dev/badge/skills/nvidia-tao/tao-skill-bank/tao-train-single-step.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 | $0.00107 | $0.01095 |
| Opus 5 | $0.00053 | $0.00548 |
| Sonnet 5 | $0.00021 | $0.00219 |
| Haiku 4.5 | $0.00011 | $0.00110 |
Grade A, and why
tao-train-single-step 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 — 85 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Normal Train
Standalone install? If this session was not initialized by the TAO skill bank plugin, run the
tao-setupskill first (host preflight, credentials, cross-skill discovery).
Standard supervised fine-tuning: train a model on a labeled dataset, optionally evaluate, then optionally export. The most common TAO workflow for adapting a pretrained model to a new dataset.
Steps
- train — executed through AutoML when the selected model has
automl_enabled: trueandautoml_policyison; setautoml_policy=offfor a plain single training run - eval — executed if
eval_dataset_uriis resolved - export — optional, on user request after training
Prerequisites
The selected model skill's resolved container_image is the default training
runtime. Do not replace it with a host venv, uv environment, generic training
image, or hand-written trainer unless the user explicitly requests that
execution mode. SDK/controller Python environments are control-plane-only; the
model action remains container-backed.
Required
- model: A compatible TAO model (e.g., clip, nvdinov2, grounding_dino)
- train_dataset_uri: URI of the training dataset (e.g.,
s3://bucket/train/) - platform: Discover the execution platforms from the installed platform skills (tao-run-on-docker / -slurm / -kubernetes / -brev, plus any external one); on a runtime that surfaces only the core router skills, read
skills/platform/tao-run-on-*/SKILL.mdfrontmatter. - container image confirmation: resolve the default image from the selected
model/action config, show it to the user, and require confirmation or
image=<override>before creating runner files or submitting training.
Optional
- eval_dataset_uri: Some model skills mark this as required — check the resolved model skill before treating it as optional.
- base_checkpoint: If not provided, defaults to the NGC pretrained checkpoint listed in the model skill, or trains from scratch if no NGC checkpoint exists.
- automl_policy:
onby default; setoffto bypass model-level AutoML for this run while leaving model metadata unchanged. Use onlyon/offin new launch settings. - image override: Use
image=<override>to pin a specific TAO toolkit build after reviewing the resolved default.
What ships with it
5 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
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 · 85 lines · 107 tokens per session scan A e55d1208fa8c
tao-train-single-step is a skill published in the GitHub repository NVIDIA-TAO/tao-skill-bank (88 stars, last pushed today), licensed Apache-2.0. It adds 107 tokens to every session and 1,095 once invoked, about $0.0005 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
nanogpt
Educational GPT implementation in 300 lines. Reproduces GPT-2 (124M) on OpenWebText. Clean, hackable code for learning transformers. By Andrej Karpathy. Perfect for understanding GPT architecture from scratch. Train on Shakespeare (CPU) or OpenWebText (multi-GPU).
ml-training-recipes
Battle-tested PyTorch training recipes for all domains — LLMs, vision, diffusion, medical imaging, protein/drug discovery, spatial omics, genomics. Covers training loops, optimizer selection (AdamW, Muon), LR scheduling, mixed precision, debugging, and systematic experimentation. Use when training or fine-tuning…
huawei-cloud-cloudrobo-model-workflow
Model development orchestration Skill covering asset query, model training, inference deployment, and real-robot evaluation in any combination. Supports full end-to-end pipeline or partial stages (e.g., train+deploy only, deploy+eval only). When user requirements involve two or more stages, prefer this Skill over…
nanogpt
Educational GPT implementation in 300 lines. Reproduces GPT-2 (124M) on OpenWebText. Clean, hackable code for learning transformers. By Andrej Karpathy. Perfect for understanding GPT architecture from scratch. Train on Shakespeare (CPU) or OpenWebText (multi-GPU).
hugging-face-model-trainer
This skill should be used when users want to train or fine-tune language models using TRL (Transformer Reinforcement Learning) on Hugging Face Jobs infrastructure. Covers SFT, DPO, GRPO and reward modeling training methods, plus GGUF conversion for local deployment. Includes guidance on the TRL Jobs package, UV…
Training
Guides fine-tuning and post-training work with explicit data, objective, hardware, and rollback assumptions.