Tao Finetune Anomalygennext

NVIDIAOfficialCreated by NVIDIA Corporation

Skill
Workflow
Data and AI
Apache-2.0
Medium risk
Has scripts

From the author

Validate a user AnomalyGenNext dataset and recipe, freeze its validation set, and prepare a canonical texture fine-tuning recipe. Use when AnomalyGenNext task weights do not yet exist. Do not use for ordinary defect generation.

— the description by NVIDIA Corporation

AutoGineer summary

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

This skill provides a structured workflow for fine-tuning AnomalyGenNext models on custom datasets. It ensures data integrity by validating dataset structures, image formats, and binary masks before training. The process includes converting base checkpoints to the required DCP format, generating a canonical training recipe, and executing the fine-tuning process within the specified AnomalyGenNext 1.1 container environment.

What it does

Validates user datasets, including defect specifications and image/mask pairings.
Converts Cosmos3-Nano checkpoints into the required DCP format.
Generates a canonical fine-tuning recipe from user templates.
Enforces strict validation of binary masks and image file extensions.
Executes LoRA fine-tuning using the AnomalyGenNext 1.1 runtime.
Verifies training success through iteration-zero and scored checkpoint validation.

When to use it

When you need to fine-tune an AnomalyGenNext model on a custom texture and defect dataset.
When AnomalyGenNext task weights do not yet exist for your specific use case.
When you need to ensure your dataset and recipe meet the strict input requirements of the AnomalyGenNext 1.1 container.
When you need to convert a Hugging Face format checkpoint for use in the AnomalyGenNext training pipeline.

Advantages

Ensures compatibility with the AnomalyGenNext 1.1 runtime through automated validation.
Prevents training failures by checking mask and image validity before GPU resources are consumed.
Standardizes the training process with canonical recipes and verified checkpoint layouts.
Provides a clear path for model acceptance based on objective performance metrics.

Usage examples

“Prepare a fine-tuning recipe for my new texture dataset located at /data/my_dataset.”: The agent runs the recipe preparation script, validates the dataset structure and masks, and outputs a canonical recipe file ready for training.
“Convert my Hugging Face Cosmos3-Nano checkpoint for use with AnomalyGenNext.”: The agent executes the conversion script to transform the checkpoint into the required DCP format, ensuring it is ready for the fine-tuning workflow.
“Train an AnomalyGenNext model using my validated canonical recipe.”: The agent launches the training workflow, mounts the necessary checkpoints and data, and monitors the process until a successful handoff is achieved.

Install

Download the skill as a ZIP:
https://autogineer.com/agent-skills/nvidia/tao-finetune-anomalygennext/download

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

mkdir -p .agents/skills && unzip tao-finetune-anomalygennext.zip -d .agents/skills

Documentation by NVIDIA Corporation

Fine-tune AnomalyGenNext

This leaf prepares a user-owned dataset and optional recipe for task-specific AnomalyGenNext LoRA training. The pinned image is declared in references/skill_info.yaml; do not replace it with the older 1.0 release. Read references/input-contract.md when validating a dataset or adapting a recipe, and references/container-runtime.md before submission.

Inputs

DATASET/
  defect_spec.jsonl
  TEXTURE/
    clean_image/*
    anomaly_image/DEFECT/*
    mask/DEFECT/*
VALIDATION/testcase.jsonl

Also supply a DCP-format Cosmos3-Nano base checkpoint directory containing checkpoint.json, model/.metadata, and at least one model/*.distcp shard, Wan2.2_VAE.pth, and a complete checkpoint root containing the required Qwen tokenizer model assets under hf/ plus facebook/dinov2-large/ for validation. The hf/ directory name is part of the upstream image's fixed checkpoint layout. The validation JSONL must contain image_filename, mask_filename, and anomaly_type; each trained TEXTURE+DEFECT needs at least three rows. A separately stored defect spec is accepted with --defect-spec.

If the official Cosmos3-Nano checkpoint is not already staged in DCP format, run the converter shipped in the pinned AnomalyGenNext image and write its output to a mounted, persistent directory:

python -m cosmos_framework.scripts.convert_model_to_dcp \
  -o /models/Cosmos3-Nano-dcp \
  --checkpoint-path Cosmos3-Nano

The registered Cosmos3-Nano name resolves the official checkpoint. To convert an already-staged Hugging Face checkpoint instead, replace it with that directory's absolute container path. Use the converter output as --base-checkpoint; preparation verifies checkpoint.json, model/.metadata, and the generated model/*.distcp shards before GPU work.

Dataset and validation images and masks must use .jpg, .jpeg, or .png, matching the extensions supported by the AnomalyGenNext 1.1 runtime loader. Every mask pixel must be binary 0 or 255; palette or grayscale masks with intermediate values are rejected before training. Each mask must contain both values, so empty/all-zero and full/all-255 masks are also rejected.

An optional user recipe.yaml is a template. Custom training settings remain, while dataset, checkpoint, validation, type order, and iteration-zero validation are replaced by validated identities. Without a template, the packaged recipe uses the established 5000-step defaults.

Prepare

After the common launch review, run the prepare_recipe action:

scripts/prepare_finetune_recipe.py \
  --dataset-root /data/my_dataset \
  --validation-testcase /data/validation/testcase.jsonl \
  --base-checkpoint /models/Cosmos3-Nano \
  --vae-path /models/Wan2.2_VAE.pth \
  --checkpoint-root /models/anomalygen-checkpoints \
  --dataset-name my_dataset \
  --recipe-template /data/recipe.yaml \
  --output /results/canonical_recipe.yaml

The action freezes absolute validation paths, validates the DCP base-checkpoint shape, rejects image extensions the runtime cannot decode, validates anomaly/mask pairing and binary masks, checks type agreement with defect_spec.jsonl, refuses output reuse, and emits a recipe plus metadata. validation_iter must be a multiple of save_iter; max_iter must reach a post-baseline validation.

Train and accept

Invoke tao-launch-workflow, review the platform, image, mounts, GPU shape, runtime, and exact recipe, then submit the train action. Bind the entire checkpoint root read-only at the fixed container path declared in skill_info.yaml; the upstream trainer resolves the Qwen tokenizer model assets at checkpoints/hf and DINOv2 at checkpoints/facebook/dinov2-large beneath its image repository. Mounting only the DINOv2 directory or setting HF_HOME does not satisfy that contract.

scripts/finetune_anomalygennext.sh \
  --recipe /results/canonical_recipe.yaml \
  --results-dir /new/training_results \
  --checkpoint-root /workspace/paidf-anomalygen/checkpoints \
  --num-gpus 1

The wrapper uses the upstream trainer already inside the pinned container. It refuses existing publication outputs and, unless --resume is explicit, an existing runtime directory. Accept success only when training_handoff.json is COMPLETE: iteration zero and a later scored checkpoint must exist, best_checkpoint.txt must select the maximum Average.nn_score, and the score must improve strictly. The handoff hashes the copied adapter and recipe.

Read references/nn-score-contract.md before changing metric or acceptance settings. A 1000-step run is the smallest quality smoke; shorter runs prove wiring only.

Files

ATTRIBUTION.mdAdded by AutoGineer951 B
BENCHMARK.md6.9 KiB
LICENSELicence file added by AutoGineer11.2 KiB
SKILL.md5.1 KiB
assets/default_recipe.yaml593 B
config/skillspector-baseline.yaml1.6 KiB
evals/evals.json1.0 KiB
references/container-runtime.md2.3 KiB
references/input-contract.md2.8 KiB
references/nn-score-contract.md837 B
references/skill_info.yaml2.5 KiB
scripts/finetune_anomalygennext.shScript3.4 KiB
scripts/prepare_finetune_recipe.pyScript12.8 KiB
scripts/tests/test_nn_improvement.pyScript3.1 KiB
scripts/tests/test_prepare_finetune_recipe.pyScript8.9 KiB
scripts/validate_nn_improvement.pyScript6.4 KiB
skill-card.md5.2 KiB
skill.oms.sig7.2 KiB

BENCHMARK.md

Skill Benchmark: tao-finetune-anomalygennext

✅ Overall verdict: PASS — Recommended for publication

Publication Recommendation

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

Evaluation Metadata

Skill: tao-finetune-anomalygennext
Evaluation date: 2026-10-16
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:bc0c5f8ff3f720d3a2430566d6a3a31037ffc1c850e46685a4204d649970a7e3 (skill-evaluator-dataset-snapshot/1)
Attempts per task: 1
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	93.2% — baseline ran, but no comparable score was available; uplift unavailable	57.7% — 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)	80.0% → 100.0% (+20.0 points)
Discoverability	95.0% — baseline ran, but no comparable score was available; uplift unavailable	20.0% — baseline ran, but no comparable score was available; uplift unavailable
Effectiveness	0.0% → 75.0% (+75.0 points)	27.5% → 27.5% (±0.0 points)
Efficiency	96.0% — baseline ran, but no comparable score was available; uplift unavailable	41.1% — 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	142,228	186,766	-44,538	-23.85%	skill 1/1; base 1/1
claude-code	prepare-anomalygennext-finetune-recipe	142,228	186,766	-44,538	-23.85%	skill 1/1; base 1/1
codex	All cases	54,183	44,465	+9,718	+21.86%	skill 1/1; base 1/1
codex	prepare-anomalygennext-finetune-recipe	54,183	44,465	+9,718	+21.86%	skill 1/1; base 1/1
ALL AGENTS	Dataset aggregate	196,411	231,231	-34,820	-15.06%	skill 2/2; base 2/2

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); 34 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: No documented scripts in table format (skills/data/tao-finetune-anomalygennext/SKILL.md)
MEDIUM QUALITY/quality_correctness: Instructions don't mention 'run_script' (skills/data/tao-finetune-anomalygennext/SKILL.md)
MEDIUM QUALITY/quality_correctness: SKILL_SPEC recommended field missing: 'metadata.tags' (skills/data/tao-finetune-anomalygennext/SKILL.md)
MEDIUM SCHEMA/frontmatter_field_placement: Root field 'tags' is ignored; use 'metadata.tags' (skills/data/tao-finetune-anomalygennext/SKILL.md)
MEDIUM SCHEMA/folder_hierarchy: Unexpected nesting depth for general skill (skills/data/tao-finetune-anomalygennext)
29 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

The author says: Requires the AnomalyGenNext 1.1 container, Pillow, and its model checkpoints.

Security

What our scanner noticed· Medium risk

Large base64-encoded blobs
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-finetune-anomalygennext · commit fce3cca

Licence

Apache License 2.0 (redistribution allowed)

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

Synced

2026-11-09

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-finetune-anomalygennext.zip