Tao Data Io

NVIDIAOfficialCreated by NVIDIA Corporation

Skill
Workflow
Automation and integrations
Apache-2.0
Medium risk
Has scripts

From the author

The data-mover for TAO jobs — decides the storage tier (A pre-positioned mount with zero fetch / B volume-from-S3 / C ephemeral in-compute fetch), stages inputs (bulk + annotation-selective + archive extract + HF/NGC PTM), maps credentials to env, routes outputs 3-way with upload-excludes, and runs the compute-frame verify gate. A support skill other platform skills (docker, kubernetes, slurm, brev, virtualenv) call to get data to and from the compute container without the TAO SDK. Trigger phrases include "stage inputs", "mount the dataset", "upload TAO results", "download only referenced files", "resolve results_dir", "verify the container can read the data".

— the description by NVIDIA Corporation

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

tao-data-io is a support skill designed to handle data movement for TAO jobs by abstracting storage strategies and staging processes. It determines the optimal storage tier—ranging from pre-positioned mounts to ephemeral in-compute fetches—and manages the transfer of datasets, annotations, and pre-trained models. The skill ensures that compute containers have the necessary data access before execution and handles the routing of results back to storage, all while maintaining strict credential security and verifying data readability from the compute environment.

What it does

Determines the storage strategy based on cluster capabilities (pre-positioned, volume-based, or ephemeral).
Stages input data including bulk files, selective annotations, archives, and pre-trained models.
Maps environment-based credentials for secure access to S3, Hugging Face, and NGC.
Verifies that the compute environment can successfully read the staged data before job launch.
Routes output results to designated storage locations or persistent mounts.
Performs selective downloads based on annotation references to optimize data transfer.

When to use it

When you need to stage datasets or pre-trained models for a TAO job.
When you need to route job outputs to S3 or a persistent mount.
When you need to verify data accessibility within a compute container before starting a job.
When you want to perform selective data downloads based on specific annotation files.

Advantages

Reduces data movement overhead by supporting pre-positioned storage strategies.
Enhances security by managing credentials via environment variables without writing to disk.
Provides a unified interface for various storage backends and data sources.
Ensures job reliability through pre-launch readability verification.
Optimizes bandwidth usage with annotation-selective download capabilities.

Usage examples

“stage inputs for my TAO job”: The agent evaluates the storage tier, fetches the required datasets or models from the configured source, and prepares the local paths for the compute container.
“verify the container can read the data”: The agent probes the compute environment to ensure the specified paths are accessible and that the container has the necessary permissions to read the input data.
“upload TAO results”: The agent identifies the output directory, applies necessary upload excludes, and synchronizes the results to the configured S3 bucket.
“download only referenced files from the annotation”: The agent executes the selective download script to fetch only the specific files listed in the provided annotation file, saving time and storage space.

Install

Download the skill as a ZIP:
https://mp-staging.autogineer.com/agent-skills/nvidia/tao-data-io/download

Unzip it into the cross-client skills folder of your project:

mkdir -p .agents/skills && unzip tao-data-io.zip -d .agents/skills

Documentation by NVIDIA Corporation

tao-data-io

Get data to and from the compute container. Decide the storage tier first — under strategy A (pre-positioned mount) no bytes move at all — and when a fetch is needed, move it host-side with aws/s5cmd/boto3/huggingface-cli/ngc directly — no nvidia-tao-sdk, no in-container runtime. Other platform skills call this skill to stage inputs before launch and sync outputs after. It never launches a container itself. The chosen tier is stamped into the job-record at submit.

When NOT to invoke this skill: if the inputs are already readable from the compute frame (a local path on the execution host, an existing Lustre/PVC/bind mount), that IS tier A — record it and skip this skill entirely; there is nothing to move. Air-gapped hosts: tier A is the only tier — never attempt an S3/HF/NGC fetch; anything missing (datasets, checkpoints, and the container images themselves) must be pre-positioned by the operator, and the preflight's readability check is the only data step that runs.

Credentials (env vars; values never written to disk by this skill)

S3 credentials use the officially documented AWS env vars, read from the session environment: AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, and (for S3-compatible stores) AWS_ENDPOINT_URL, AWS_DEFAULT_REGION. The aws CLI and boto3 pick the variables up natively — never run aws configure and never write ~/.aws/credentials:

set -a; source /path/to/.env; set +a   # omit if already exported
aws s3 ls "s3://$S3_BUCKET_NAME/..."   # reads AWS_* from the environment

If a session provides only the legacy TAO names (ACCESS_KEY, SECRET_KEY, S3_ENDPOINT_URL, CLOUD_REGION), map them once, scoped to the command: AWS_ACCESS_KEY_ID="$ACCESS_KEY" AWS_SECRET_ACCESS_KEY="$SECRET_KEY" aws s3 ...

HF_TOKEN / NGC_KEY pass through unchanged for PTM pulls. Never pass a credential as a CLI argument (-p, --token, -e KEY=value); use --password-stdin or -e VAR (no value).

Storage strategy (agent decides; the gate verifies)

Pick per backend from what the cluster/daemon actually offers:

A — pre-positioned mount (PVC / NFS / Lustre / bind mount already holding the data): mount it, author the mount paths into the spec. No S3 fetch. Also the air-gap answer.
B — volume populated from S3: an initContainer / stage step fills a durable volume; compute mounts it; a final step drains it back to S3 if needed.
C — ephemeral + in-compute fetch: no persistent mount — fetch into the compute container and upload out at the end (today's K8s default).

Verify-before-launch gate (the one invariant)

The path the spec references is readable from the compute frame (not the launcher's), and the output destination persists after the container exits.

Probe in the compute's frame of reference (in-container aws s3 cp/touch on the resolved results_dir, or a kubectl run/srun probe) — a green aws s3 ls on the launcher is not proof the pod can read the data (managed backends inject different creds into the compute container).

Input staging

Bulk folder: s5cmd cp 's3://.../*' <stage> or aws s3 sync.
Single file: aws s3 cp.
Annotation-selective (download only files referenced by an annotation): use references/selective_download.py (below).
Archive: tar -xzf X -C <dir> --strip-components=1 guarded by a .extracted marker (idempotent).
PTM (ngc:// / hf://): huggingface-cli download / ngc registry model download-version, then author the local path.

After staging, author the spec with local paths and run the verify gate.

Output routing (3-way) + upload

TAO_RESULTS_ROOT set → write to that mount, no upload.
else S3_BUCKET_NAME set → upload to s3://$S3_BUCKET_NAME/results/$TAO_JOB_ID/.
else → loud ephemeral warning.

SLURM: never set S3_BUCKET_NAME (Lustre-only); run any upload on the login node, not inside the GPU allocation. Upload with excludes: aws s3 sync <local>/ s3://... --exclude '.tao/*' <upload_excludes...>.

Quick Start — annotation-selective staging

Download only the files an annotation references (e.g. the video column), preserving relative paths, into a local staging dir:

set -a; source /path/to/.env; set +a   # omit if already exported
python references/selective_download.py \
  --annotation /path/to/annotation.parquet \
  --key video \
  --bucket "$S3_BUCKET_NAME" --src-prefix datasets/clips \
  --dest /data/stage/clips

--key is repeatable or comma-separated; --format overrides extension inference (parquet/jsonl/json/csv).

Helpers

references/selective_download.py — annotation-driven selective download (boto3 + pandas). Unit tests live in references/tests/; run with python -m pytest.

Files

ATTRIBUTION.mdAdded by AutoGineer903 B
BENCHMARK.md6.8 KiB
LICENSELicence file added by AutoGineer11.2 KiB
SKILL.md5.9 KiB
config/skillspector-baseline.yaml1.6 KiB
evals/evals.json826 B
references/selective_download.pyScript7.5 KiB
references/tests/test_selective_download.pyScript6.8 KiB
skill-card.md4.1 KiB
skill.oms.sig5.2 KiB

BENCHMARK.md

Skill Benchmark: tao-data-io

✅ Overall verdict: PASS — Recommended for publication

Publication Recommendation

Recommended for publication based on the completed evaluation evidence in this report.

Evaluation Metadata

Skill: tao-data-io
Evaluation date: 2026-09-17
Evaluator version: 1.5.6
Agents: Claude Code (aws/anthropic/bedrock-claude-opus-4-8), Codex (openai/openai/gpt-5.5)
Tasks: 1 evaluation tasks (1 positive)
Dataset digest: sha256:9df227152dd4d53d020412fb78f24217f966c8a4553452c2910a01ae6d87b881 (skill-evaluator-dataset-snapshot/1)
Attempts per task: 3
Environment: k8s-sandbox
Tier 2 evidence: required for publication
Tier 3 evidence: required for publication

Each task attempt ran in its own isolated sandbox pod.

What This Report Answers

The three-tier evaluation checks whether the skill:

is safe to use;
produces correct answers;
is discovered and activated when needed;
helps the agent complete the user's goal and expected workflow; and
avoids wasted skill and tool usage.

Results at a Glance

Measure	Claude Code (Baseline → Skill Uplift)	Codex (Baseline → Skill Uplift)
Overall	92.7% — baseline ran, but no comparable score was available; uplift unavailable	95.4% — baseline ran, but no comparable score was available; uplift unavailable
Security	100.0% → 100.0% (±0.0 points)	100.0% → 100.0% (±0.0 points)
Correctness	0.0% → 100.0% (+100.0 points)	0.0% → 100.0% (+100.0 points)
Discoverability	100.0% — baseline ran, but no comparable score was available; uplift unavailable	95.0% — baseline ran, but no comparable score was available; uplift unavailable
Effectiveness	16.7% → 100.0% (+83.3 points)	16.7% → 83.3% (+66.6 points)
Efficiency	63.4% — baseline ran, but no comparable score was available; uplift unavailable	98.5% — baseline ran, but no comparable score was available; uplift unavailable

How to read this table: baseline is the same task attempted without the target skill. Scores are rounded to one decimal; threshold-adjacent values use additional precision so their displayed band matches the verdict. Uplift is derived from those displayed scores and shown in percentage points.

Example: 47.0% → 92.0% (+45.0 points) means the skill-assisted run scored 92.0%, 45.0 percentage points above its 47.0% no-skill baseline.

Token Usage

Actual Tier 3 execution usage is reported for every observed agent/case pair and both conditions.

Agent	Dataset case	With skill	Without skill	Delta	Change	Coverage
claude-code	All cases	105,471	90,096	N/A	N/A	skill 1/1; base 3/3
claude-code	tao-data-io-staging	105,471	90,096	N/A	N/A	skill 1/1; base 3/3
codex	All cases	29,813	41,035	N/A	N/A	skill 1/1; base 3/3
codex	tao-data-io-staging	29,813	41,035	N/A	N/A	skill 1/1; base 3/3
ALL AGENTS	Dataset aggregate	135,284	131,131	N/A	N/A	skill 2/2; base 6/6

Prompt tokens include cached reads, so total tokens are prompt + completion (cached is not added twice). The Efficiency score uses (prompt - cached) + completion. N/A means the relevant trajectory counters were not available; coverage is never estimated.

Tier Status

Tier	Purpose	Status	Evidence
Tier 1	Static validation	PASSED WITH OBSERVATIONS	11 validator(s); 19 finding(s)
Tier 2	Semantic deduplication	PASSED	2 validator(s); 0 finding(s)
Tier 3	Live agent evaluation	PASS	2 agent(s); 1 task(s)

Findings and Observations

<details> <summary>Show detailed findings and successful checks</summary>

MEDIUM QUALITY/quality_correctness: SKILL_SPEC recommended field missing: 'metadata.tags' (skills/platform/tao-data-io/SKILL.md)
MEDIUM SCHEMA/frontmatter_field_placement: Root field 'tags' is ignored; use 'metadata.tags' (skills/platform/tao-data-io/SKILL.md)
MEDIUM SCHEMA/folder_hierarchy: Unexpected nesting depth for general skill (skills/platform/tao-data-io)
MEDIUM SCHEMA/body_recommended_section: Missing recommended section: '## Instructions' (skills/platform/tao-data-io/SKILL.md)
MEDIUM SCHEMA/body_recommended_section: Missing recommended section: '## Examples' (skills/platform/tao-data-io/SKILL.md)
14 additional finding(s) are available in the full evaluation artifacts.

</details>

Scoring Methodology

<details> <summary>Show dimension definitions, source signals, and thresholds</summary>

Dimension	Question	Scored signals
Security	Is it safe to use?	`security` (100%)
Correctness	Is the answer correct?	`accuracy` (100%)
Discoverability	Was the right skill loaded when needed?	`skill_execution` (100%)
Effectiveness	Did the skill help complete the task?	`goal_accuracy` (50%) + `behavior_check` (50%)
Efficiency	Did it avoid wasted tool calls and token usage?	`skill_efficiency` (50%) + `token_efficiency` (50%)

Dimension bands: PASS at 50% or above; NEUTRAL from 40% to below 50%; FAIL below 40%.
Overall Tier 3 lift: PASS at +5 points or more; FAIL at -10 points or less; values between those bands are NEUTRAL.
Overall verdict: PASS only when every configured dimension passes for at least one supported agent. Lift is reported as diagnostic evidence and does not override this gate.
The 50% attempt pass threshold is a separate per-task gate; it is not the dimension pass threshold.
Effectiveness is the equal-weight mean of goal completion (goal_accuracy) and expected workflow adherence (behavior_check).
Efficiency is 50% tool-call productivity (the backward-compatible skill_efficiency wire id) and 50% token_efficiency. Positive-case skill routing is scored under Discoverability, not Efficiency; a negative case without a routing target is N/A. N/A sources are omitted, remaining weights are renormalized, and the dimension is marked partial.

Signals present in this run:

security (Security): unsafe operations, secret leakage, and unauthorized access.
skill_execution (Skill Execution): whether the expected skill was selected, decoys were avoided, and the workflow executed.
skill_efficiency (Tool Productivity): tool-call productivity (legacy wire id; routing is scored under Discoverability).
accuracy (Accuracy): final-answer correctness against the reference answer.
goal_accuracy (Goal Accuracy): whether the user's goal was achieved.
behavior_check (Behavior Check): whether the expected workflow behavior was followed.
token_efficiency (Token Efficiency): actual uncached prompt plus completion usage (50% of Efficiency).

</details>

Freshness

Regenerate this benchmark when the skill, evaluation dataset, target agent/model, evaluator version, environment, or scoring policy changes.

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Declares allowed tools, Needs Python, Needs Node.js

The author says: Requires aws CLI or s5cmd on the staging host, plus Python 3.10+ with boto3 and pandas/pyarrow for annotation-selective download. No nvidia-tao-sdk, no fsspec/s3fs. Credentials are read from the process environment, whether exported in the user's shell or sourced from a user-approved env file.

Security

What our scanner noticed· Medium risk

Large base64-encoded blobs
References to credential files
Scripts that your agent may run

This skill includes scripts that your agent may run. Read them before you use it.

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by

NVIDIA Corporation

Publisher

NVIDIA (Official)

Source

NVIDIA/skills · skills/tao-data-io · commit fce3cca

Licence

Apache License 2.0 (redistribution allowed)

LICENSE (added by AutoGineer from LICENSE-APACHE)
ATTRIBUTION.md (written by AutoGineer)

Synced

2026-11-08

AutoGineer is not affiliated with or endorsed by NVIDIA Corporation or NVIDIA.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download tao-data-io.zip