CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

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. Also covers LLaMA-Factory and Unsloth (guides in references/).

Quality

81%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Passed

No findings from the security scan

SKILL.md
Quality
Evals
Security

Quality

Content

75%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 well-organized, highly actionable reference: executable code for every common case, a logical setup-to-deployment sequence, and clearly signaled reference files. The main improvements are tightening marketing-flavored asides, fixing the few snippets with undefined variables or missing imports, and pushing the method/integration/benchmark catalogs out to the already-present reference files.

Suggestions

Fix the small executability gaps in the code examples: add `import torch` to the standard LoRA quick start (the data_collator lambda uses torch.stack), and define or annotate the `inputs`/`prompts` placeholders in the multi-adapter serving and vLLM snippets.

Move content that belongs at reference depth out of SKILL.md — the PEFT methods comparison with IA3/Prefix Tuning code, the Axolotl/vLLM integration patterns, and the performance benchmark tables could live in references/advanced-usage.md or a new references/benchmarks.md, keeping SKILL.md as a tighter overview.

Flatten the nested reference structure in the merged sub-skills: references/llama-factory/overview.md and references/unsloth/overview.md each point into a second-level references/ directory, so surface the key files directly in the SKILL.md table or restructure to keep references one level deep.

DimensionReasoningScore

Conciseness

The body is dense and code/table-driven with almost no conceptual padding — comments carry real information ("# Rank (8-64, higher = more capacity)", "# NormalFloat4 (best for LLMs)"). Minor trimmable instances remain: marketing asides ("70B model now fits on single 24GB GPU!", "Trains only 0.01% of parameters!") and inline catalogs (full method comparison, three integration frameworks, benchmark tables) that duplicate what Claude partly knows or could live in references.

4 / 5

Actionability

Nearly all guidance is copy-paste-ready executable code covering the common cases: full LoRA pipeline, QLoRA config, adapter loading/merging, multi-adapter switching, IA3, Prefix Tuning, TRL, Axolotl YAML, and vLLM. Small gaps prevent a 5: the quick-start's data_collator uses torch.stack without importing torch, and the multi-adapter/vLLM snippets reference undefined variables (inputs, prompts).

4 / 5

Workflow Clarity

A clear progression from "When to use PEFT" through installation, quick start, parameter selection, loading/merging, and deployment, with a "Common issues" section giving ranked recovery options (OOM solutions 1/2/3) and a best-practice checkpoint ("Evaluate on held-out data before merging"). Not a 5 because validation checkpoints are implicit rather than explicit steps in the main fine-tuning workflow.

4 / 5

Progressive Disclosure

Good structure: quick start inline, advanced material clearly signaled one level deep via an annotated References section ("[Advanced Usage](references/advanced-usage.md) - DoRA, LoftQ...", "[Troubleshooting](references/troubleshooting.md)") and a Related-tools table pointing to the llama-factory/unsloth overviews, all of which exist on disk. Held at 4 because the ~430-line body inlines sizable catalogs (method comparison, three integration patterns, benchmark tables) that belong in the existing reference files, and the llama-factory/unsloth sub-guides themselves nest a second references/ level.

4 / 5

Total

16

/

20

Passed

Description

83%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 description with an explicit "Use when..." clause carrying concrete trigger conditions (GPU memory limits, <1% parameters, multi-adapter serving) and specific method names. Voice is appropriately third-person and the niche is well-differentiated; only minor gaps in keyword synonyms and action enumeration keep it from top marks.

DimensionReasoningScore

Specificity

Lists several concrete capabilities — "fine-tuning large models (7B-70B)", "using LoRA, QLoRA, and 25+ methods", "train <1% of parameters", "multi-adapter serving" — with specific numbers and method names. Not a 5 because "25+ methods" is unenumerated and the HuggingFace/transformers sentence is ecosystem context rather than a distinct action.

4 / 5

Completeness

Explicitly answers both: what ("Parameter-efficient fine-tuning for LLMs using LoRA, QLoRA, and 25+ methods... HuggingFace's official library") and when ("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") with concrete trigger conditions.

5 / 5

Trigger Term Quality

Strong natural terms users would actually say: "fine-tuning", "LoRA", "QLoRA", "limited GPU memory", "multi-adapter serving", "7B-70B". A few common variations are missing (e.g. "adapters", "memory-efficient fine-tuning", "fine-tune on a consumer GPU"), keeping it below the comprehensive-synonym coverage of a 5.

4 / 5

Distinctiveness Conflict Risk

The PEFT/LoRA/QLoRA niche is clear with distinct technical triggers, and the description even scopes adjacent tools (LLaMA-Factory, Unsloth) as sub-guides. Minor overlap risk remains with generic fine-tuning or HuggingFace training skills for the broad phrase "fine-tuning large models".

4 / 5

Total

17

/

20

Passed

Validation

81%

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 13 / 16 Passed

Validation for skill structure

CriteriaDescriptionResult

frontmatter_unknown_keys

Unknown frontmatter key(s) found; consider removing or moving to metadata

Warning

relative_links

Relative link issues: 2 deeper-than-1-level

Warning

referenced_paths_exist

Referenced path issues: 4 deeper-than-1-level

Warning

Total

13

/

16

Passed

Repository

OpenLAIR/dr-claw

Path

skills/fine-tuning/peft/SKILL.md

Reviewed

Table of Contents

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.