What does this skill do, and when should you use it?
This skill guides NVIDIA TAO Toolkit workflows for training, exporting, and running inference with an NVDINOv2 vision-transformer backbone. It uses teacher-student self-distillation without labels and produces general-purpose visual features. Training is AutoML-enabled and can be routed through AutoML or direct training per run. The workflow targets Docker and NVIDIA GPU environments, with the default ViT-Large setup requiring substantial memory.
Reads the packaged TAO Core JSON schemas, YAML spec templates, AutoML configuration, and deployment reference; constructs required train and inference data-source paths; runs TAO training, AutoML, export, and inference workflows; hands selected student_epoch_*.pth checkpoints to export and inference; produces an ONNX file and documents the TAO Deploy TensorRT flow.
- A computer-vision researcher performs ViT self-supervised pretraining without labeled data.
- A model engineer generates general-purpose visual features for downstream tasks.
- A TAO user uses AutoML to select training configurations and the best model.
- A deployment engineer exports a selected NVDINOv2 checkpoint to ONNX or generates a TensorRT engine.
- A platform engineer runs Lightning-managed training across multiple GPUs or nodes.
How do you install this skill?
- Before execution, confirm Docker, NVIDIA Container Toolkit, the TAO container, A100 40GB+ GPUs, and all data sources are available.
- Reconcile checkpoint fields between skill_info.yaml and the inference template, and verify that student_epoch_*.pth—not a full training checkpoint—is supplied where required.
- Obtain user confirmation before handling datasets, checkpoints, or encryption_key, and avoid placing sensitive values in logs or external services.
- The benchmark covers only one positive planning task and does not establish successful real training, export, or inference.
- Shell / CLI
- Network access
- Local filesystem
DockerNVIDIA Container ToolkitA100 GPUs with 40GB+ VRAM
Install the skill with the Skills CLI command documented in the repository README:
npx skills add nvidia/skills --skill tao-train-nvdinov2 --yes
The runtime also requires Docker, NVIDIA Container Toolkit, and NVIDIA GPUs meeting the memory guidance. The README does not document more detailed TAO container or version installation steps.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/tao-train-nvdinov2" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- self-supervised ViT pretraining,
- visual representation learning.
After installation, ask the Agent to “train NVDINOv2,” “self-supervised ViT pretraining,” “DINOv2 backbone,” or “visual representation learning.” Before a training-stage request, read references/skill_info.yaml and resolve automl_policy. Training and inference require the corresponding images_train.tar.gz or images_test.tar.gz data-source path. For export and inference, use the selected student_epoch_*.pth checkpoint; before gen_trt_engine, read references/tao-deploy-nvdinov2.md.
What are this skill's strengths and limitations?
- Supports label-free teacher-student self-distillation.
- Covers training, AutoML, export, inference, and TAO Deploy workflows.
- Includes schema, template, checkpoint handoff, and common-error guidance.
- Supports Lightning-managed multi-GPU and multi-node training.
- The default ViT-Large teacher and student are highly memory-intensive; the skill requires at least 40GB-class GPUs and strongly recommends 4–8 GPUs.
- Requires Docker and NVIDIA Container Toolkit.
- SSL training does not use labels, and evaluation is downstream-task-specific.
- The source declares version 0.1.0; its default 10 epochs are intended for quick testing, while production runs typically use 100+ epochs.
- PyT inference does not accept a TensorRT engine, and no standalone distill action is available.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| TAO NVDINOv2 Visual Representation Training this page ✓ NVIDIA · Official | 47 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO Metric-Learning Recognition ✓ NVIDIA · Official | 49 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO FastFoundationStereo Real-Time Stereo Depth ✓ NVIDIA · Official | 55 · Use with care | ★ 3.8k | today | Apache-2.0 |
| OCDNet Scene Text Detection ✓ NVIDIA · Official | 53 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO FoundationStereo Stereo Depth ✓ NVIDIA · Official | 52 · Use with care | ★ 3.8k | today | Apache-2.0 |
How did FollowSkills review this skill?
The skill limits declared tools to Read and Bash and explicitly requires Docker/NVIDIA Container Toolkit; no malware, credential theft, or covert exfiltration is shown. However, training, export, and inference access datasets, checkpoints, result directories, and an encryption key, without user-confirmation guidance, data-flow disclosure, sensitive-data handling, rollback, or dependency-integrity details, so points are deducted.
The documentation covers AutoML routing, data sources, checkpoint selection, and several error patterns, making the main path plausible. Static evidence still shows configuration-consistency risk around checkpoint fields between skill_info.yaml and the inference template; there is no reproducible key-path test coverage or comprehensive abnormal-input feedback. The static ceiling also limits this dimension to 10, so points are deducted.
Trigger phrases, train/export/inference scenarios, data format, hardware requirements, and the unavailable standalone distill action are described. The skill primarily fits users with A100 40GB+ GPUs, Docker, and TAO; Chinese-language support, non-fit boundaries, and network reachability are not addressed, limiting applicability and causing deductions.
The skill has a name, description, Apache-2.0 license, version, author, parameter tables, templates, error patterns, and repository-level synchronization context. It lacks the recommended Instructions and Examples sections, uses an incomplete author format, and provides no changelog, explicit maintainer/update path, or systematic FAQ. The repository benchmark also records schema and layout findings, so points are deducted.
The documentation can guide construction of training, export, and inference configurations and explains checkpoint handoff, the AutoML metric, and common failures, so the core value is clear. No execution occurred in this static review, and there is insufficient representative output evidence that results are directly usable; field inconsistencies may require rework. The static ceiling therefore requires a conservative deduction.
Versioned templates, skill_info.yaml, an evaluation task, and a benchmark report provide limited audit evidence. The evaluation contains only one positive task and no negative tasks, and supplies no independently checkable logs or committed tests covering the skill’s key execution paths. Under the static ceiling, points are deducted.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →