tao-finetune-clip

CLIP vision-language model for image-text retrieval, zero-shot classification, embedding extraction, ONNX export, and TensorRT deployment. Use when fine-tuning or training CLIP, running zero-shot classification, computing image embeddings, or deploying CLIP to ONNX/TensorRT.

Quality

84%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Low

Low-risk findings worth noting

Quality

Content

76%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

A highly actionable, well-organized body with concrete spec blocks and commands for every action, weakened mainly by absent validation feedback loops for batch/destructive operations and several references pointing to files that do not exist in the bundle.

Suggestions

Add explicit validation checkpoints to the train and deploy workflows (e.g., after training, assert val/t2i_mAP improved and checkpoints exist; after gen_trt_engine, verify the engine loads before running TensorRT evaluate/inference) to lift workflow_clarity above the batch-operation cap.

Fix broken references: replace references/model_info.yaml with the actual references/skill_info.yaml, remove or create defaults.json / clip.config.json / schemas/train.schema.json (the schemas/ directory does not exist), and clarify the spec_template_deploy_* prefix since only spec_template_deploy.yaml is present.

Disambiguate the line-38 "config.json for action/data-source wiring" reference, since the root config.json is the skill-scoring config, not an action/data-source wiring file — this misleads at runtime.

Dimension	Reasoning	Score
Conciseness	The body is information-dense and assumes Claude's competence (no padding about what ONNX/TensorRT/CLIP are), with nearly every token being model-specific operational detail Claude would not know; it stops short of a 5 only because the checkpoint-handling discussion recurs in a couple of places and the overall volume could be slightly tightened.	4 / 5
Actionability	Copy-paste-ready spec override blocks for every action (train/evaluate/inference/export/gen_trt_engine), exact commands ("clip gen_trt_engine -e {config_path}", "tao deploy clip gen_trt_engine -e /path/to/spec.yaml"), concrete env vars (TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1), and exact model IDs make this fully executable across the common cases.	5 / 5
Workflow Clarity	A clear sequence exists (identify action → load files → collect datasets → place in spec_overrides → run in correct image), but training and TensorRT deployment are batch/destructive operations with only implicit validation (val/t2i_mAP, "increase only if validation loss is still improving") and no explicit validate→fix→retry checkpoints, so the batch-operation cap at 3 applies.	3 / 5
Progressive Disclosure	Section structure is strong and references are clearly signaled one level deep (skill_info.yaml, spec_template.yaml, error-patterns.md, spec-param-inference.md, tao-deploy-clip.md), but several cited paths are broken (defaults.json, references/model_info.yaml, schemas/train.schema.json, clip.config.json, and the spec_template_deploy_* prefix that resolves to a single file), which is a moderate organization gap keeping it off the top anchor.	4 / 5
	Total	16 / 20 Passed

Description

92%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

A strong, well-structured description that clearly states both capabilities and trigger conditions with concrete, natural language. The only minor gap is a lack of synonym variation in trigger terms.

Dimension	Reasoning	Score
Specificity	Lists five concrete actions — "image-text retrieval, zero-shot classification, embedding extraction, ONNX export, and TensorRT deployment" — giving comprehensive coverage of CLIP capabilities, matching the multiple-specific-actions anchor.	5 / 5
Completeness	Explicitly answers both what ("CLIP vision-language model for image-text retrieval, zero-shot classification, embedding extraction, ONNX export, and TensorRT deployment") and when ("Use when fine-tuning or training CLIP...") with concrete trigger phrases, matching the top anchor.	5 / 5
Trigger Term Quality	Natural terms like "fine-tuning or training CLIP", "running zero-shot classification", "computing image embeddings", and "deploying CLIP to ONNX/TensorRT" are well covered, but few synonyms/variations are included and CLIP has no characteristic file extensions, so it sits just below comprehensive.	4 / 5
Distinctiveness Conflict Risk	The CLIP-specific niche and CLIP-tied triggers ("training CLIP", "deploying CLIP to ONNX/TensorRT") create a clear niche with minimal overlap risk against other skills.	5 / 5
	Total	19 / 20 Passed

Validation

87%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 14 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning
referenced_paths_exist	Referenced path issues: 1 missing	Warning

	Total	14 / 16 Passed

Repository: NVIDIA/skills
Path: skills/tao-finetune-clip/SKILL.md
Commit: 20bb6aa

Reviewed: 3 days ago

Table of Contents

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.