model-pruning

Reduce LLM size and accelerate inference using pruning techniques like Wanda and SparseGPT. Use when compressing models without retraining, achieving 50% sparsity with minimal accuracy loss, or enabling faster inference on hardware accelerators. Covers unstructured pruning, structured pruning, N:M sparsity, magnitude pruning, and one-shot methods.

Quality

75%

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

Fix and improve this skill with Tessl

tessl review fix ./19-emerging-techniques/model-pruning/SKILL.md

Quality

Content

50%

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

A content-rich, code-heavy skill that is genuinely actionable in its headline examples but suffers from internal redundancy, undefined helper placeholders in the strategies/pipeline code, and a lack of validation checkpoints in the destructive pruning workflow. The unused wanda.md reference indicates progressive-disclosure structure that was not wired up.

Suggestions

Link and defer to references/wanda.md for the deep Wanda treatment, and move Core Concepts / Strategies / Performance Comparison into reference files so the body becomes a lean overview with one-level-deep, clearly signaled links.

Replace or define the placeholder helpers (load_calibration_data, prune_layer, train_step, fine_tune, prune_model) and fix the non-executable 'importance = weight^2 / diag(Hessian)' so every code block is copy-paste runnable, and add the missing `from datasets import load_dataset` import.

Add an explicit validation checkpoint to the pruning pipeline (e.g., verify sparsity level and run a quick perplexity/eval probe on calibration data before saving) so the destructive weight-mutation step has a validate-then-proceed feedback loop.

Dimension	Reasoning	Score
Conciseness	The ~485-line body is mostly code and not padded with concepts Claude already knows, but it repeats the same technique across Quick Start, Core Concepts, and Best Practices (e.g., the Wanda \|weight\|×activation criterion appears three times) and could be tightened; not the 'lean, every token earns its place' anchor.	2 / 3
Actionability	The centerpiece Wanda and N:M snippets are executable and copy-paste ready, but several examples rely on undefined placeholders — load_calibration_data(), prune_layer(), train_step(), fine_tune(), prune_model(), the un-imported load_dataset(), and the non-Python 'importance = weight^2 / diag(Hessian)' — so key details are missing.	2 / 3
Workflow Clarity	The production pipeline gives a numbered 1–5 sequence, but pruning mutates model weights (a destructive/batch-style operation) with no mid-workflow validation checkpoint before saving; the post-hoc Evaluation comparison does not satisfy the explicit validate-then-proceed feedback loop the top anchor requires.	2 / 3
Progressive Disclosure	Sections provide structure, but the file is a monolithic ~485-line wall with Core Concepts, Strategies, and Performance Comparison inlined that could be split out, and the existing references/wanda.md bundle file is never linked or signaled from the body.	2 / 3
	Total	8 / 12 Passed

Description

100%

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, third-person description that states concrete capabilities, provides an explicit 'Use when' trigger with multiple natural conditions, and carves out a distinctive niche. Voice is correctly third person, so no specificity penalty applies.

Dimension	Reasoning	Score
Specificity	Lists multiple concrete actions — 'Reduce LLM size', 'accelerate inference', 'compressing models without retraining', 'achieving 50% sparsity' — plus an enumerated coverage list (unstructured, structured, N:M, magnitude, one-shot), matching the 'lists multiple specific concrete actions' anchor.	3 / 3
Completeness	Clearly answers both 'what' (reduce LLM size / accelerate inference via pruning techniques) and 'when' with an explicit 'Use when compressing models without retraining, achieving 50% sparsity..., or enabling faster inference...' trigger clause, satisfying the highest anchor.	3 / 3
Trigger Term Quality	Contains natural user-facing terms like 'compressing models', 'faster inference', 'pruning', 'sparsity', and 'hardware accelerators' alongside the named techniques; good coverage of phrasing a user would naturally say when they need this skill.	3 / 3
Distinctiveness Conflict Risk	Occupies a clear niche (LLM pruning/compression) anchored by distinctive named techniques (Wanda, SparseGPT, N:M sparsity) and specific triggers, making it unlikely to fire for an unrelated skill.	3 / 3
	Total	12 / 12 Passed

Validation

93%

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 — 15 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	15 / 16 Passed

Repository: Orchestra-Research/AI-Research-SKILLs
Path: 19-emerging-techniques/model-pruning/SKILL.md
Commit: 773a529

Reviewed: about 11 hours 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.