Tao Train Dinov3

What it does

When your agent uses it

• Phrases include train DINOv3
• Domain-adapt a foundation backbone
• Continual pretraining
• Self-supervised finetune DINOv3

Example prompts

• “train DINOv3”
• “DINOv3 SSL”
• “domain-adapt a foundation backbone”
• “/tao-train-dinov3”

Requirements

• Python 3
• Docker
• Compatibility (from SKILL.md): Requires docker + nvidia-container-toolkit.
• Pre-approved tools (allowed-tools): Read, Bash

Workflow steps

5 steps, taken from the first numbered list in SKILL.md.

1. 1Obtain a compatible DINOv3 checkpoint in timm or TAO format.
2. 2Point dataset.train_dataset.images_dir to a directory of unlabeled images.
3. 3Configure the backbone, training scale, and output location.
4. 4Run training and evaluate useful teacher checkpoints on data representative of the downstream task.
5. 5Use the selected teacher checkpoint with convert, export, or inference.

What it can do on your machine

Read from SKILL.md and the folder at commit dfdd080. It shows what the files ask for, not the result of running them.

• Tool permissions

  Pre-approves these tools, so the agent can use them without asking each time:

  • Read
  • Bash

  From allowed-tools in the SKILL.md frontmatter.

• Runs code

  Shell commands in SKILL.md call:

  • docker

  From the folder's file list and the shell code blocks in SKILL.md.

• Network

  No URLs in SKILL.md. Its commands use docker, which can reach the network depending on how they are called.

  From URLs in SKILL.md, links to its own repository left out.

• Credentials

  Names no API keys, tokens, secrets or passwords.

  From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.

• Compatibility

  Requires docker + nvidia-container-toolkit.

  From compatibility in the SKILL.md frontmatter.

Context cost

Tao Train Dinov3 loads about 2.2k tokens when it runs, and up to ~8.2k if it reads all its reference files. Until then it costs about 110 tokens; SKILL.md has 825 words of instructions outside code blocks.

Always · name and description, kept in context so the agent knows when to use it

~110

When it runs · the whole SKILL.md, loaded when a task matches

~2.2k

With references · SKILL.md plus every file in references/, read only if the agent opens them

~8.2k

Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.

Safety

• NotePre-approves every shell command (allowed-tools: Bash)SKILL.md

  allowed-tools: Read, Bash

Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); files beside SKILL.md are not scanned.

SKILL.md

The full file from NVIDIA/skills at commit dfdd080, republished under its Apache-2.0 licence (© NVIDIA). 825 words, ~2,207 tokens.

name

tao-train-dinov3

description

DINOv3 continual self-supervised pre-training. Domain-adapts public DINOv3 ViT backbones on unlabeled images via teacher-student self-distillation (DINO + iBOT + KoLeo, optional Gram anchoring) and converts the EMA teacher into a timm-format backbone for downstream tasks. Trigger phrases include "train DINOv3", "DINOv3 SSL", "domain-adapt a foundation backbone", "continual pretraining", "self-supervised finetune DINOv3".

allowed-tools

Read, Bash

compatibility

Requires docker + nvidia-container-toolkit.

license

Apache-2.0

metadata.version

0.1.0

metadata.author

NVIDIA Corporation

tags

self, supervised, learning, dinov3

TAO_PYT_IMAGE_DEFAULT=nvcr.io/nvidia/tao/tao-toolkit:7.2.0-pyt  # versions-key: images.tao_toolkit.pyt
TAO_PYT_IMAGE="${TAO_PYT_IMAGE:-$TAO_PYT_IMAGE_DEFAULT}"
RUN_ROOT="${RUN_ROOT:-$PWD}"
DOCKER_COMMON=(
  --rm --gpus all --shm-size=8g
  --shm-size=8g
  --ulimit memlock=-1
  --ulimit stack=67108864
  -v "$RUN_ROOT/data:/data:ro"
  -v "$RUN_ROOT/specs:/specs:ro"
  -v "$RUN_ROOT/results:/results"
)

docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  dinov3 train -e /specs/train.yaml

docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  dinov3 inference -e /specs/inference.yaml

docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  dinov3 export -e /specs/export.yaml

docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  dinov3 convert -e /specs/convert.yaml

model:
  backbone:
    teacher_type: vit_b
    student_type: vit_b
    rope_theta: 100.0
dataset:
  train_dataset:
    images_dir: /path/to/unlabeled/images
  batch_size: 16
train:
  pretrained_model_path: /path/to/dinov3/weights
  num_epochs: 10
  num_gpus: 1

© NVIDIA, Apache-2.0. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file

Files

SKILL.md and 18 other files (references) in skills/tao-train-dinov3 of NVIDIA/skills.

• SKILL.md
• BENCHMARK.md
• config/skillspector-baseline.yaml
• evals/evals.json
• references/dinov3-method.md
• references/dinov3-recipes.md
• references/skill_info.yaml
• references/spec_template_convert.yaml
• references/spec_template_export.yaml
• references/spec_template_inference.yaml
• references/spec_template_train.yaml
• references/spec_template_train_highres.yaml
• schemas/convert.schema.json
• schemas/export.schema.json
• schemas/inference.schema.json
• schemas/manifest.json
• schemas/train.schema.json
• … and 2 more

Open the folder on GitHubat commit dfdd080