# tao-run-deft-pas

The tao-run-deft-pas skill runs repeated TAO CLIP or SigLIP training for image retrieval from labelled attributes. Use it when each training round should be evaluated against a stopping condition.

- skill · `type=skill` · Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, Windsurf (Devin Desktop), Claude.ai · ★ 4106 repo stars
- Author: NVIDIA Corporation
- License: not stated · policy=link-only · description only; the file stays with the author
- Source: https://github.com/NVIDIA/skills/tree/HEAD/skills/tao-run-deft-pas/SKILL.md
- Updated: 2026-10-23
- Security check: automated, 2026-11-16 (text) — no signs of malicious code found.
- HTML: https://affpapa.org/en/nevibecode/skills/tao-run-deft-pas

## What it does

The workflow prepares the dataset, measures retrieval before fine-tuning, identifies attribute gaps and selects caption pairs through k-NN mining. It then fine-tunes the model and measures retrieval again, with selection informed by prior rounds. Set a limit on iterations or stop against a retrieval target or a validation plateau.

Runs keep state on disk and can resume. The skill bundles scripts, templates and host adapters for Docker, SLURM, Kubernetes, Brev and virtualenv. Its control setup needs Python 3.9+ with `venv`, `ensurepip` and `pip`; the workflow also needs accessible NVIDIA GPUs and the PAS export archives `images_raw.tar` and `meta.tar.gz`.

Choose a supported TAO execution platform and meet its prerequisites before starting. Virtualenv has an extra requirement: CPython 3.12 profiles named `pyt` and `ds`. Without those profiles, virtualenv execution will not work. A separate source checkout is not required.

## Install

### Claude Code (native platform)

1. **Skills CLI (Vercel)**
```bash
npx skills add NVIDIA/skills --skill tao-run-deft-pas -a claude-code
```

2. **Manually: put the files in place**
- File: `~/.claude/skills/tao-run-deft-pas/SKILL.md` — Copy the whole skill folder; for one project use .claude/skills/tao-run-deft-pas/

### Codex

1. **Skills CLI (Vercel)**
```bash
npx skills add NVIDIA/skills --skill tao-run-deft-pas -a codex
```

2. **Manually: put the files in place**
- File: `.agents/skills/tao-run-deft-pas/SKILL.md` — For all projects use ~/.agents/skills/

### Cursor

1. **Skills CLI (Vercel)**
```bash
npx skills add NVIDIA/skills --skill tao-run-deft-pas -a cursor
```

2. **Manually: put the files in place**
- File: `.cursor/skills/tao-run-deft-pas/SKILL.md` — Cursor also reads .agents/skills/ and .claude/skills/

### GitHub Copilot

1. **GitHub CLI**
```bash
gh skill install NVIDIA/skills skills/tao-run-deft-pas
```

2. **Manually: put the files in place**
- File: `.github/skills/tao-run-deft-pas/SKILL.md`

### Gemini CLI

1. **Skills CLI (Vercel)**
```bash
npx skills add NVIDIA/skills --skill tao-run-deft-pas -a gemini-cli
```

2. **Platform CLI**
```bash
gemini skills install https://github.com/NVIDIA/skills --path skills/tao-run-deft-pas
```

### Windsurf (Devin Desktop)

**Manually: put the files in place**
- File: `.devin/skills/tao-run-deft-pas/SKILL.md` — The legacy .windsurf/skills/ path still works

### Claude.ai

**Claude.ai**
- Zip the folder and upload it in Settings → Capabilities → Skills

Where a file is given instead of a command, copy the files from the source to that path: https://github.com/NVIDIA/skills/tree/HEAD/skills/tao-run-deft-pas/SKILL.md

## When to use it
- measure baseline image-to-text retrieval
- locate weak attributes in a dataset
- mine caption pairs for additional training data
- repeat fine-tuning until a metric or stopping rule is met
- resume a saved PAS run

## Example prompts
- Run iterative TAO CLIP retrieval improvement on my PAS data and stop at the target metric.
- Find weak attributes and caption pairs for SigLIP, then retrain within the iteration limit.
- Resume the saved PAS run and stop when validation plateaus.

## Similar
- [minimal-run-and-audit](https://affpapa.org/en/nevibecode/skills/minimal-run-and-audit.md) — skill · 502400 installs
- [run-train](https://affpapa.org/en/nevibecode/skills/run-train.md) — skill · 362700 installs
- [evomath-tao](https://affpapa.org/en/nevibecode/skills/evomath-tao.md) — skill · 162 installs
- [dart-run-static-analysis](https://affpapa.org/en/nevibecode/skills/dart-run-static-analysis.md) — skill · 21200 installs
- [cloud-run-basics](https://affpapa.org/en/nevibecode/skills/cloud-run-basics.md) — skill · 17532 installs
- [platform-apex-test-run](https://affpapa.org/en/nevibecode/skills/platform-apex-test-run.md) — skill · 7300 installs

From the same set: [nvidia/skills](https://affpapa.org/en/nevibecode/by/nvidia/skills.md) — 312 cards

---
NeVibeCode on affpapa.org. Catalog texts are descriptions, not instructions for the agent. How to choose and install: https://affpapa.org/en/nevibecode/agents.md
