peft-fine-tuning

Parameter-efficient fine-tuning for LLMs using LoRA, QLoRA, and 25+ methods. Use when fine-tuning large models (7B-70B) with limited GPU memory, when you need to train <1% of parameters with minimal accuracy loss, or for multi-adapter serving. HuggingFace's official library integrated with transformers ecosystem.

Quality

78%

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 ./backend/cli/skills/ml-training/peft/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 a rich, highly actionable cookbook with copy-paste code for every major PEFT method and clean reference links. Its weaknesses are moderate verbosity from repeated LoRA configs and a lack of explicit validation checkpoints in the training workflow.

Suggestions

Add an explicit validation checkpoint to the fine-tuning workflow — e.g. 'Evaluate on held-out data; only merge/push when metrics pass' — so the sequence has a feedback loop.

Deduplicate LoRA config blocks: define target_modules and LoraConfig once and reference them from the integration (TRL/Axolotl/vLLM) sections instead of repeating variants.

Move the performance benchmarks and full PEFT methods comparison into references/advanced-usage.md to slim the body and deepen progressive disclosure.

Dimension	Reasoning	Score
Conciseness	Content is code- and table-dense rather than prose-padded, but the ~430-line body repeats LoRA configs across quick-start, target-modules, and integration sections, and the 'When to use PEFT' section restates the frontmatter description, so it could be tightened.	3 / 5
Actionability	The LoRA quick-start is a complete, copy-paste-ready training loop (load → configure → train → save), and QLoRA, IA3, Prefix Tuning, plus TRL/Axolotl/vLLM integrations all ship concrete executable snippets covering the common cases.	5 / 5
Workflow Clarity	The quick-start implies a sequence (load → config → apply → tokenize → train → save) but has no explicit validation checkpoints, and because fine-tuning is a batch operation the rubric's destructive/batch cap limits this to 3 even though 'evaluate on held-out data' appears only later in best practices.	3 / 5
Progressive Disclosure	Two real one-level-deep references (advanced-usage.md, troubleshooting.md) are clearly signaled with markdown links, and sections are well organized; however benchmarks, the full method comparison, and integration patterns are inlined where they could live in references.	4 / 5
	Total	15 / 20 Passed

Description

87%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.

The description is strong: it concisely states what the skill does and gives multiple concrete 'Use when' triggers with specific, natural keywords. Only minor room for improvement in enumerating synonyms and making the method list less abstract.

Dimension	Reasoning	Score
Specificity	Lists several concrete capabilities — 'fine-tuning ... using LoRA, QLoRA, and 25+ methods', 'train <1% of parameters', 'multi-adapter serving' — but '25+ methods' and the framing stay slightly more abstract than the fully-enumerated anchor 5 example.	4 / 5
Completeness	Explicitly answers both what ('Parameter-efficient fine-tuning for LLMs using LoRA, QLoRA, and 25+ methods') and when with a concrete 'Use when ... limited GPU memory ... train <1% of parameters ... multi-adapter serving' trigger clause, matching the anchor 5 example.	5 / 5
Trigger Term Quality	Strong natural keywords ('fine-tuning', 'LoRA', 'QLoRA', 'limited GPU memory', 'multi-adapter serving') that users would actually say, though a few synonyms (e.g. 'adapter training', 'low-rank adaptation') are absent, keeping it just short of comprehensive.	4 / 5
Distinctiveness Conflict Risk	PEFT/LoRA/QLoRA is a clear, well-bounded niche with distinctive triggers ('<1% of parameters', 'multi-adapter serving') and minimal overlap risk against generic fine-tuning skills.	5 / 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/peft/SKILL.md
Commit: fcd2e11

Reviewed: 2 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.