gptq

Post-training 4-bit quantization for LLMs with minimal accuracy loss. Use for deploying large models (70B, 405B) on consumer GPUs, when you need 4× memory reduction with <2% perplexity degradation, or for faster inference (3-4× speedup) vs FP16. Integrates with transformers and PEFT for QLoRA fine-tuning.

Quality

80%

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

72%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-structured skill with excellent progressive disclosure and copy-paste code across all major use cases. The main gaps are explanatory padding Claude doesn't need and the absence of explicit post-quantization validation checkpoints.

Suggestions

Trim the 'How GPTQ works' conceptual steps and the group-counting worked example, which restate knowledge Claude already has; keep only the group-size trade-off table.

Add an explicit validation checkpoint after model.quantize() — e.g., a perplexity check against the FP16 baseline or a generation sanity test — so the quantization workflow has a feedback loop for the destructive weight modification.

Reorder the quantize-your-own-model section into a numbered sequence with a 'Verify' step between quantize and save/push.

Dimension	Reasoning	Score
Conciseness	Mostly efficient with a code-first approach, but sections like 'How GPTQ works' and the worked group-counting example explain concepts (group-wise quantization mechanics) Claude already knows and could be trimmed.	3 / 5
Actionability	Fully executable, copy-paste-ready code spans installation, loading, quantizing, kernel backends, transformers/PEFT integration, QLoRA, and multi-GPU deployment, covering the common cases concretely.	5 / 5
Workflow Clarity	The quantize-your-own-model flow (load → config → calibrate → quantize → save) is sequenced, but there are no explicit validation checkpoints (e.g., perplexity check or output-equivalence check after quantization); per rubric guidance, missing validation for a batch/destructive operation caps this at 3.	3 / 5
Progressive Disclosure	SKILL.md is a clear overview with well-signaled, one-level-deep references to real bundle files (calibration.md, integration.md, troubleshooting.md all exist), with content appropriately split and easy to navigate.	5 / 5
	Total	16 / 20 Passed

Description

88%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, specific description that clearly conveys both the skill's purpose and its triggering conditions with concrete metrics. Minor improvements could add a few more natural synonyms (e.g., 'weight compression', 'quantize a model') to broaden trigger coverage.

Dimension	Reasoning	Score
Specificity	Lists multiple concrete actions — 'Post-training 4-bit quantization', '4× memory reduction', 'faster inference (3-4× speedup)', and 'QLoRA fine-tuning' — with comprehensive coverage of the skill's capabilities.	5 / 5
Completeness	Explicitly answers both 'what' (post-training 4-bit quantization for LLMs with minimal accuracy loss) and 'when' ('Use for deploying large models... when you need 4× memory reduction... or for faster inference').	5 / 5
Trigger Term Quality	Includes natural user phrases like '4-bit quantization', 'deploying large models on consumer GPUs', and 'QLoRA fine-tuning', but is missing some common synonyms and variations users might say.	4 / 5
Distinctiveness Conflict Risk	The GPTQ 4-bit post-training quantization niche is clearly distinct with specific triggers, though it has minor overlap risk with adjacent quantization methods like AWQ and bitsandbytes.	4 / 5
	Total	18 / 20 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: synthetic-sciences/openscience
Path: backend/cli/skills/ml-training/gptq/SKILL.md
Commit: 3a6c3a9

Reviewed: 1 day 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.