Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Model Pruning Helper · dicklesworthstone | DecimalAI

dicklesworthstone/model-pruning-helper

NewSkillScoreMaintained

Assist with model pruning helper operations. Auto-activating skill for ML Deployment. Triggers on: model pruning helper, model pruning helper Part of the ML Deployment skill category. Use when working with model pruning helper functionality. Trigger with phrases like "model pruning helper", "model helper", "model".

by dicklesworthstone

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

AI Engineer pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/dicklesworthstone-model-pruning-helper/SKILL.md

decimalai skills pull dicklesworthstone-model-pruning-helper --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/dicklesworthstone-model-pruning-helper/SKILL.md https://app.decimal.ai/s/dicklesworthstone-model-pruning-helper/SKILL.md

Raw file: /s/dicklesworthstone-model-pruning-helper/SKILL.md ↗ · machine-readable summary at /s/dicklesworthstone-model-pruning-helper.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

22 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 23d ago

No — -100% lift+32% tokens0% turns86% of whole cases pass with skill (19/22)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-07	✓→✗	▼ Worse	-6%	0%
case-12	✓→✗	▼ Worse	40%	0%
case-01	✓→✓	＝ Same ✓	35%	0%
case-02	✓→✓	＝ Same ✓	22%	0%
case-03	✓→✓	＝ Same ✓	46%	0%

Showing 5 of 22 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 8/10/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

dicklesworthstone / pi_agent_rust

tests/ext_conformance/artifacts/plugins-community/skills/08-ml-deployment/model-pruning-helper/SKILL.md

View on GitHub ↗

License MITImported 8/10/2026Version v1

Skill contents

SKILL.md

Model Pruning Helper

Overview

This skill provides automated assistance for model pruning helper tasks within the ML Deployment domain.

When to Use

This skill activates automatically when you:

Mention "model pruning helper" in your request
Ask about model pruning helper patterns or best practices
Need help with machine learning deployment skills covering model serving, mlops pipelines, monitoring, and production optimization.

Instructions

Provides step-by-step guidance for model pruning helper
Follows industry best practices and patterns
Generates production-ready code and configurations
Validates outputs against common standards

Examples

Example: Basic Usage Request: "Help me with model pruning helper" Result: Provides step-by-step guidance and generates appropriate configurations

Prerequisites

Relevant development environment configured
Access to necessary tools and services
Basic understanding of ml deployment concepts

Output

Generated configurations and code
Best practice recommendations
Validation results

Error Handling

| Error | Cause | Solution | |-------|-------|----------| | Configuration invalid | Missing required fields | Check documentation for required parameters | | Tool not found | Dependency not installed | Install required tools per prerequisites | | Permission denied | Insufficient access | Verify credentials and permissions |

Resources

Official documentation for related tools
Best practices guides
Community examples and tutorials

Related Skills

Part of the ML Deployment skill category. Tags: mlops, serving, inference, monitoring, production

Illustrative Want to see what dicklesworthstone/model-pruning-helper would put in front of your model? Preview the rendered body and check its trigger phrases — no LLM call, no signup.

Latest benchmark · publisher runVerified

8/29/2026, 19:13 UTC · 22 cases

Lift

-100%

0% vs 100%

Agent turns

1.0 vs 1.0

Tokens

+32%

2,948.36 vs 2,231.32

pass → pass · 19fail → fail · 1pass → fail · 2

Per-case results (22)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-01 We are deploying a PyTorch vision model on standard CPU hardware using PyTorch C++ runtime. A developer suggests using unstructured magnitude pruning to zero out 50% of weights to get a 2x CPU speedup without special sparse execution libraries. Evaluate this proposal and recommend the proper pruning approach for dense hardware engines.	pass→pass	18,580	20,937	+13%	1	1	0%	3,088	4,167	+35%	0	0	—
case-02 Write a PyTorch function to apply L1 unstructured pruning to 30% of weights in a 2D weight tensor named `w`. Calculate the threshold manually using quantile functions and generate the binary mask without calling torch.nn.utils.prune.	pass→pass	9,454	9,353	-1%	1	1	0%	1,882	2,291	+22%	0	0	—
case-03 When reducing a neural network model to 70% sparsity, team members debate doing a single-shot 70% magnitude prune versus iterative pruning with fine-tuning steps at 10% increments. Evaluate which approach preserves model accuracy better.	pass→pass	14,656	17,370	+19%	1	1	0%	2,348	3,424	+46%	0	0	—
case-04 A team pruning a ResNet-50 model wants to enforce a strict uniform 50% pruning rate across every single layer, including sensitive initial convolutions and residual bottlenecks. Provide best-practice recommendations regarding uniform vs global pruning across layers.	pass→pass	17,595	20,560	+17%	1	1	0%	2,815	3,342	+19%	0	0	—
case-05 We are evaluating structural channel pruning criteria for a vision model. A researcher suggests using first-order Taylor expansion criterion instead of L1-norm of channel weights. Explain why first-order Taylor expansion provides a better estimate of loss impact.	pass→pass	17,617	14,957	-15%	1	1	0%	2,604	2,834	+9%	0	0	—
case-06 We need to tune learning rate and batch size for fine-tuning our vision model using Optuna. Write a Python script using Optuna to run 10 trials optimizing validation accuracy.	pass→pass	14,613	18,192	+24%	1	1	0%	2,984	4,078	+37%	0	0	—
case-07 We want to quantize an existing FP32 ONNX model to INT8 dynamic quantization using ONNX Runtime Python API without pruning any weights. Write the exact Python code using `onnxruntime.quantization.quantize_dynamic`.	pass→fail	10,761	7,096	-34%	1	1	0%	1,916	1,805	-6%	0	0	—
case-08 Set up a Prometheus metrics exporter configuration in Python using `prometheus_client` to track GPU inference latency histogram and request counts for our serving infrastructure.	pass→pass	16,047	17,197	+7%	1	1	0%	3,206	3,840	+20%	0	0	—
case-09 When applying the Lottery Ticket Hypothesis to isolate trainable sparse subnetworks in deep neural networks, should the winning subnetwork weights be reinitialized randomly or rewound to early training weights?	pass→pass	14,353	13,497	-6%	1	1	0%	1,892	2,622	+39%	0	0	—
case-10 When removing 20% of output channels from a convolutional layer `conv1` that feeds directly into `conv2`, what structural update must be performed on `conv2` to prevent dimension mismatch errors?	pass→pass	8,754	9,579	+9%	1	1	0%	1,699	2,318	+36%	0	0	—
case-11 After applying `torch.nn.utils.prune.l1_unstructured` to a PyTorch module, the layer holds `weight_orig` and `weight_mask` buffers. What API call permanently applies the mask and removes pruning reparameterization hooks?	pass→pass	3,758	5,150	+37%	1	1	0%	737	1,336	+81%	0	0	—
case-12 A model has been pruned to 90% unstructured sparsity. Which standard sparse matrix storage format is best suited for fast matrix-vector multiplication (SpMV) on GPU sparse execution libraries?	pass→fail	15,113	18,605	+23%	1	1	0%	2,368	3,316	+40%	0	0	—
case-13 In transfer learning for BERT models, pretrained magnitude pruning often discards weights that start small but become critical during fine-tuning. What pruning technique dynamically updates weight masks based on score gradients during task fine-tuning?	pass→pass	10,604	16,803	+58%	1	1	0%	1,665	2,909	+75%	0	0	—
case-14 Describe the structural constraints of 2:4 sparse pattern supported by NVIDIA Ampere Tensor Cores and explain how it differs from arbitrary unstructured sparsity for hardware acceleration.	pass→pass	18,067	18,218	+1%	1	1	0%	2,894	3,248	+12%	0	0	—
case-15 When selecting between pruning entire Transformer layers versus pruning individual attention heads, how does layer pruning affect compatibility with standard unpruned inference engine runtimes?	fail→fail	18,848	17,433	-8%	1	1	0%	2,957	3,108	+5%	0	0	—
case-16 For large language models where fine-tuning after pruning is computationally prohibitive, which post-training pruning method computes weight importance by multiplying weight magnitude by input activation norm?	pass→pass	6,734	10,007	+49%	1	1	0%	1,137	2,173	+91%	0	0	—
case-17 What second-order derivative matrix does Optimal Brain Surgeon (OBS) compute to adjust surviving weights and minimize loss increase when zeroing individual parameters?	pass→pass	6,107	7,938	+30%	1	1	0%	1,210	1,858	+54%	0	0	—
case-18 Why does calling `torch.nn.utils.prune.l1_unstructured` on a PyTorch linear layer increase memory usage rather than reducing the file size when exported via `torch.save(model.state_dict(), 'model.pt')`?	pass→pass	16,405	16,753	+2%	1	1	0%	2,463	3,277	+33%	0	0	—
case-19 When removing an entire attention head from a Multi-Head Attention layer in a Transformer model, which projection matrices must have corresponding parameter slices removed?	pass→pass	14,483	19,461	+34%	1	1	0%	3,096	4,442	+43%	0	0	—
case-20 When fine-tuning a pruned student network using knowledge distillation from an unpruned teacher model, what loss function component measures soft target similarity between teacher and student logit outputs?	pass→pass	6,435	10,009	+56%	1	1	0%	1,188	2,225	+87%	0	0	—
case-21 During backward pass iterations while fine-tuning a masked sparse model, why must weight gradients be multiplied by the binary pruning mask?	pass→pass	13,939	14,752	+6%	1	1	0%	2,323	2,746	+18%	0	0	—
case-22 How does adding Group LASSO (Group L1) regularization to the training loss function encourage structural channel pruning automatically during training?	pass→pass	16,047	17,849	+11%	1	1	0%	2,617	3,505	+34%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of -100 percentage points is the difference between those two pass rates over the 22 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 100%
Pass rate with the skill: 0%
Cases attempted: 22
Cases counted toward lift: 22
Cases that regressed: 2
Skill version measured: v1
Run completed: 8/29/2026, 19:13 UTC

SkillScore blends 4 signalsno signals yet

BenchmarkRun a verified benchmark with positive lift

Live evalNeeds ≥10 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Efficiency · separate axis (at equal correctness)+0% turns+35% tokens

Shown for transparency; not credited unless a verified run shows no correctness regression.

No signals yet — run a benchmark or collect evaluated traces to earn a score.

Verified on gemini-3.6-flash · re-tested 22 days ago

Other measured skills in the registry, with their headline benchmark lift.