quantizing-models-bitsandbytes

Quantizes LLMs to 8-bit or 4-bit for 50-75% memory reduction with minimal accuracy loss. Use when GPU memory is limited, need to fit larger models, or want faster inference. Supports INT8, NF4, FP4 formats, QLoRA training, and 8-bit optimizers. Works with HuggingFace Transformers.

Quality

78%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Fix and improve this skill with Tessl

tessl review fix ./skills/optimization/bitsandbytes/SKILL.md

Quality

Content

68%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.

The body is highly actionable with complete, executable code and well-organized sections that push advanced detail to genuine reference files. Its weaknesses are redundancy between the Quick start and the first workflow, and missing validation/feedback checkpoints in the QLoRA and 8-bit-optimizer workflows.

Suggestions

De-duplicate the 8-bit/4-bit config blocks: keep the detailed config in Workflow 1 Step 3 and have the Quick start reference it, rather than repeating both configs verbatim.

Add explicit validation/feedback steps to the QLoRA and 8-bit-optimizer workflows (e.g., verify trainable params, run a sanity forward pass, confirm memory savings before continuing).

Trim the inline QLoRA training detail now that references/qlora-training.md exists, keeping the body as an overview that points to the reference for the full fine-tuning procedure.

Dimension	Reasoning	Score
Conciseness	Mostly concrete code and tables, but the 8-bit/4-bit config blocks and memory numbers are repeated between the Quick start and Workflow 1 Step 3, and QLoRA material appears both inline and in references/qlora-training.md, so it could be tightened.	3 / 5
Actionability	Fully executable, copy-paste-ready code throughout — BitsAndBytesConfig setups, from_pretrained calls, LoRA config, and 8-bit optimizer usage all cover the common cases concretely.	5 / 5
Workflow Clarity	Each workflow has a sequenced checklist, but only Workflow 1 includes a verification (test inference + check memory); Workflows 2 and 3 lack explicit validation or feedback loops, leaving checkpoints implicit.	3 / 5
Progressive Disclosure	Clear section structure with a well-signaled, one-level-deep Advanced topics block pointing to three real files (qlora-training.md, quantization-formats.md, memory-optimization.md), though some advanced content is duplicated inline rather than purely referenced.	4 / 5
	Total	15 / 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, third-person description that clearly states both capability and trigger conditions with concrete, quantified outcomes. It is specific and well-scoped; the only gap is the omission of the underlying tool name (bitsandbytes) and a few synonym triggers.

Dimension	Reasoning	Score
Specificity	Lists multiple concrete actions — "Quantizes LLMs to 8-bit or 4-bit", "Supports INT8, NF4, FP4 formats, QLoRA training, and 8-bit optimizers" — giving comprehensive coverage of capabilities.	5 / 5
Completeness	Explicitly answers what ("Quantizes LLMs to 8-bit or 4-bit for 50-75% memory reduction") and when ("Use when GPU memory is limited, need to fit larger models, or want faster inference") with concrete trigger phrases.	5 / 5
Trigger Term Quality	Good natural triggers like "GPU memory is limited", "need to fit larger models", "want faster inference" plus format names, but missing some synonyms and extension-style variations that would push it to comprehensive.	4 / 5
Distinctiveness Conflict Risk	The INT8/NF4/FP4, QLoRA, and 8-bit-optimizer scope plus "Works with HuggingFace Transformers" carves a mostly-distinct niche, though the tool name itself is absent so minor overlap with other quantization skills remains.	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: OpenLAIR/dr-claw
Path: skills/optimization/bitsandbytes/SKILL.md
Commit: 4b35756

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.