
Tao run deft aoi cosmos3
Runs an iterative DEFT loop that improves Cosmos3 bare OK/NG AOI inspection accuracy
- Scripts
- shell, Python
- Platforms
- Claude Code, Codex, Cursor
- GitHub stars
- 4K
- Last commit
- 2 days ago
- License
- Apache-2.0
- Listing updated
- 1 day ago
Similar skills
About Tao run deft aoi cosmos3
Tao run deft aoi cosmos3 is an open-source agent skill for ai models and llm tooling. It is available on Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf and Goose.
This skill drives a disk-backed improvement loop for NVIDIA Cosmos Reason 3 (Cosmos3) reasoners aimed at automated optical inspection. It defaults to the Nano variant and uses Edge or Super only when explicitly requested. Each iteration evaluates the base model on a frozen Benchmark split and a Proxy split, mines real image pairs that match Proxy errors, generates synthetic NG defect pairs, assembles a training file, and trains with LoRA SFT on the Cosmos Framework.
The workflow is a state machine recorded in a state file, with validation scripts that check label format, split isolation and a frozen Benchmark hash before each stage. Work runs through the user's chosen compute platform using submit, status, logs and cancel commands, and a pre-flight summary must be approved before any launch. Labels are restricted to bare OK or NG responses; rich or reasoning annotation modes are not supported.
The skill expects ShareGPT annotation files and image folders supplied by the user, a prepared Qwen3-VL checkpoint conversion, the Cosmos Framework container image, and related TAO skills for model, anomaly generation and mining. It is installed as part of a full TAO skill bank and uses Python with pyarrow and yaml.
When your agent uses it
Run the disk-backed DEFT AOI improvement loop for NVIDIA Cosmos Reason 3 / Cosmos3 models, using Nano by default and Edge or Super when explicitly requested: evaluate the base model on Proxy and frozen Benchmark splits, mine real image pairs from Proxy gaps, generate AnomalyGen synthetic NG pairs, assemble a per-iteration Train JSON from both producers, train with Cosmos Framework LoRA SFT, and repeat through the selected platform's submit/status/logs/cancel contract. This migration supports bare labels only: the assistant response must be exactly OK or NG. Use for "run Cosmos3 DEFT AOI", "CR3 AOI loop", or "improve Cosmos3 PCB inspection with bare OK/NG"; do not use for rich/reasoning annotation, one-off Cosmos training, or generic anomaly generation.
From the skill's SKILL.md. Your agent reads this description to decide when to load the skill.
What it can access
- Runs shell and Python scripts(see the files on GitHub)
- Asks for: Read, Task, Bash, Write(see the files on GitHub)
- No network calls found in its scripts(see the files on GitHub)
- License: Apache-2.0(see the files on GitHub)
From a static check of its 61 files (609 KB) on GitHub on 17 November 2026. Nothing was run. Read the files before you install a skill: it can tell your agent to run anything its tools allow.
Check another skill before you install it.
Install
Copy the skill folder to ~/.claude/skills/tao-run-deft-aoi-cosmos3 to use it in all your projects:
tmp=$(mktemp -d) &&
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp" &&
mkdir -p ~/.claude/skills &&
rm -rf ~/.claude/skills/tao-run-deft-aoi-cosmos3 &&
cp -R "$tmp/skills/tao-run-deft-aoi-cosmos3" ~/.claude/skills/tao-run-deft-aoi-cosmos3 &&
rm -rf "$tmp"For one project only, put the folder in .claude/skills/tao-run-deft-aoi-cosmos3 inside it instead.
Copy the skill folder to ~/.agents/skills/tao-run-deft-aoi-cosmos3 to use it in every repository:
tmp=$(mktemp -d) &&
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp" &&
mkdir -p ~/.agents/skills &&
rm -rf ~/.agents/skills/tao-run-deft-aoi-cosmos3 &&
cp -R "$tmp/skills/tao-run-deft-aoi-cosmos3" ~/.agents/skills/tao-run-deft-aoi-cosmos3 &&
rm -rf "$tmp"For one repository only, put the folder in .agents/skills/tao-run-deft-aoi-cosmos3 inside it instead.
Copy the skill folder to ~/.cursor/skills/tao-run-deft-aoi-cosmos3 to use it in every project:
tmp=$(mktemp -d) &&
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp" &&
mkdir -p ~/.cursor/skills &&
rm -rf ~/.cursor/skills/tao-run-deft-aoi-cosmos3 &&
cp -R "$tmp/skills/tao-run-deft-aoi-cosmos3" ~/.cursor/skills/tao-run-deft-aoi-cosmos3 &&
rm -rf "$tmp"For one project only, put the folder in .cursor/skills/tao-run-deft-aoi-cosmos3 inside it instead.
Run in a terminal (add --scope workspace to install it for the current workspace only):
gemini skills install https://github.com/NVIDIA/skills.git --path skills/tao-run-deft-aoi-cosmos3Or copy the folder to ~/.gemini/skills/tao-run-deft-aoi-cosmos3 yourself.
For GitHub Copilot in VS Code, copy the skill folder to ~/.copilot/skills/tao-run-deft-aoi-cosmos3:
tmp=$(mktemp -d) &&
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp" &&
mkdir -p ~/.copilot/skills &&
rm -rf ~/.copilot/skills/tao-run-deft-aoi-cosmos3 &&
cp -R "$tmp/skills/tao-run-deft-aoi-cosmos3" ~/.copilot/skills/tao-run-deft-aoi-cosmos3 &&
rm -rf "$tmp"For one project only, put the folder in .github/skills/tao-run-deft-aoi-cosmos3 inside it instead.
Get the folder:
git clone --depth 1 https://github.com/NVIDIA/skills.gitThen copy skills/skills/tao-run-deft-aoi-cosmos3 into your agent's skills folder as tao-run-deft-aoi-cosmos3. Agents that support SKILL.md read one folder per skill.
What it does
- Frozen Benchmark gating
- Evaluates a frozen Benchmark split at baseline and each iteration to decide whether the loop should stop.
- Proxy error analysis
- Reviews Proxy false accepts and false rejects to produce root-cause artifacts that guide mining targets.
- Real image mining
- Selects real image pairs near Proxy gaps using a cosine floor and history checks to avoid reuse.
- Synthetic NG generation
- Creates synthetic defect pairs with the AnomalyGen tool and converts them into bare NG training records.
- LoRA training workflow
- Renders Framework training and evaluation configs, trains with LoRA SFT and resumes from prior checkpoints.
- Resumable run state
- Tracks each stage, duration and event in a state file so interrupted runs can resume and report progress.
Good for
- Improve a Cosmos3 model's PCB defect detection with bare OK/NG labels
- Run an iterative training loop until a Benchmark recall target is met
- Mine real defect image pairs from evaluation gaps for a new training set
- Generate synthetic NG defect pairs to cover missed defects
- Evaluate a Cosmos Reason 3 base model zero-shot on an inspection benchmark
Tao run deft aoi cosmos3 FAQ
Which Cosmos3 variant does it use by default?
It uses Nano by default. Edge or Super are used only when explicitly requested, and it never switches variants silently.
What label format does it produce?
Training responses must be exactly OK or NG. Rich, reasoning and multiple-choice annotation modes are not supported.
Does it start training without approval?
No. A launch review and a pre-flight summary must be explicitly approved before any state, dependency install or native launch.
What inputs are needed?
User-supplied Proxy and Benchmark annotation files, a mining pool, and images. Specs are generated by the workflow rather than supplied.
How do I install Tao run deft aoi cosmos3 in Claude Code?
Copy the skill folder from GitHub to ~/.claude/skills/tao-run-deft-aoi-cosmos3 to use it in all your projects, or to .claude/skills/tao-run-deft-aoi-cosmos3 in one project.
Does Tao run deft aoi cosmos3 run scripts?
Yes. It runs shell and Python scripts. No network calls were found in its scripts. This comes from a static check of the files on GitHub; nothing was run. Read the files before you install it.
Which AI agents can use Tao run deft aoi cosmos3?
Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf and Goose. Each reads skills in the SKILL.md format from its own folder; the install steps on this page list them.
What are the best alternatives to Tao run deft aoi cosmos3?
The closest alternatives to Tao run deft aoi cosmos3 are Tao train rtdetr, Tao run deft cr its mining, Tao train dinov3, Tao run automl deft pipeline and Tao finetune cosmos reason.
Where this comes from
- Description: from the skill's own SKILL.md.
More from NVIDIA/skills

Guides an agent through reducing peak GPU memory in Megatron Bridge training runs
- Open source
- 4K GitHub stars
- Knowledge and memory

Guides tuning of MoE expert-parallel communication overlap in Megatron Bridge
- Open source
- 4K GitHub stars
- AI models and LLM tooling

Guides choice of MoE token dispatcher (alltoall, DeepEP, HybridEP) by hardware and EP size
- Open source
- 4K GitHub stars
- AI models and LLM tooling

Hardware-specific MoE training playbooks and candidate parallelism configs for Megatron
- Open source
- 4K GitHub stars
- AI models and LLM tooling

Guides long-context MoE training in Megatron Bridge: CP sizing, recompute and dispatcher
- Open source
- 4K GitHub stars
- AI models and LLM tooling

Guides an agent through evidence-gated MoE performance tuning in Megatron Bridge
- Open source
- 4K GitHub stars
- Developer tools



