CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

fine-tuning-peft

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

77%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

Fix and improve this skill with Tessl

tessl review fix ./cli-tool/components/skills/ai-research/fine-tuning-peft/SKILL.md
SKILL.md
Quality
Evals
Security

Quality

Content

67%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 reference for PEFT with copy-paste-ready code, good decision tables, and a properly wired one-level-deep bundle. Its weaknesses are redundancy: the inline Common issues section overlaps the troubleshooting reference, and the Best practices section recaps earlier content, both of which cost token efficiency.

Suggestions

Remove or shrink the inline 'Common issues' section to a two-line pointer to references/troubleshooting.md, keeping only the most frequent error and its one-line fix, to eliminate duplication with the bundle file.

Trim the 'Best practices' list to items not already covered in the rank/alpha guidance and QLoRA sections, or fold it into those sections as inline notes.

Make the quick-start script fully self-contained (add the missing `import torch`, define or note the expected dataset variable) so the primary example is truly copy-paste ready.

DimensionReasoningScore

Conciseness

The body is mostly dense, useful material (code, tables, parameter guidance) with little explanation of concepts Claude already knows, but it could be tightened: the 'Common issues' section duplicates content that also lives in references/troubleshooting.md (CUDA OOM, adapter not applying), and the 'Best practices' list largely recaps guidance already given in the parameter-selection and QLoRA sections. This matches 'mostly efficient but includes some unnecessary explanation or could be trimmed' rather than level 4, where only minor instances of over-explanation would remain.

3 / 5

Actionability

Guidance is overwhelmingly concrete and near copy-paste ready: complete LoRA and QLoRA configs with realistic values and expected output ('trainable params: 13,631,488 || ... 0.17%'), loading/merging/multi-adapter snippets, and integration examples for TRL, Axolotl, and vLLM. It falls short of level 5 only due to minor gaps such as the quick-start script using torch in its data collator without importing it, and a few snippets referencing undefined variables (prompts, inputs, dataset).

4 / 5

Workflow Clarity

The body follows a coherent progression (install -> configure -> train -> save -> load -> merge -> serve -> debug), and validation checkpoints are present (model.print_trainable_parameters(), 'Verify adapter is active' checks, 'Evaluate on held-out data before merging'). It is not a level 5 because the sequence is implied by section order rather than explicitly staged, and some checkpoints (e.g., verifying a merge before deployment) are only mentioned in a best-practices list rather than in the merge workflow itself.

4 / 5

Progressive Disclosure

Both bundle files exist, are referenced one level deep, and are clearly signaled with content descriptions ('[Advanced Usage](references/advanced-usage.md) - DoRA, LoftQ, rank stabilization, custom modules'), and neither reference file nests further references. It sits at level 4 rather than 5 because the ~420-line SKILL.md inlines a 'Common issues' section that duplicates references/troubleshooting.md, so content is not optimally split between the overview and the bundle.

4 / 5

Total

15

/

20

Passed

Description

78%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 that explicitly states what the skill does and when to use it, with good natural trigger terms (LoRA, QLoRA, fine-tuning, GPU memory) and a well-differentiated niche. Its main weaknesses are the second-person 'when you need to' phrasing, which costs it specificity under the rubric, and the vague '25+ methods' claim with unnamed methods.

Suggestions

Rewrite the trigger clause in third person to avoid the voice penalty, e.g., 'Use when fine-tuning large models (7B-70B) with limited GPU memory, when training under 1% of parameters with minimal accuracy loss is needed, or for multi-adapter serving.'

Replace the unnamed '25+ methods' over-claim with one or two additional concrete method names users might mention (e.g., 'IA3, prefix tuning') to strengthen specificity and trigger coverage.

DimensionReasoningScore

Specificity

The description lists several concrete capabilities ('Parameter-efficient fine-tuning for LLMs using LoRA, QLoRA, and 25+ methods', 'multi-adapter serving', 'train <1% of parameters'), which on its own matches the several-specific-actions level (4), but the second-person clause 'when you need to train <1% of parameters' triggers the rubric's one-point voice penalty, bringing it to 3. It is above the names-domain-with-1-2-actions level (3 anchor without penalty) because coverage of the domain's capabilities is broad, though '25+ methods' is an unnamed over-claim.

3 / 5

Completeness

It clearly answers both questions: what ('Parameter-efficient fine-tuning for LLMs using LoRA, QLoRA, and 25+ methods') 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 explicit, concrete trigger phrases. This matches the anchor for explicit what AND when with concrete triggers.

5 / 5

Trigger Term Quality

Good natural keyword coverage: 'fine-tuning', 'LoRA', 'QLoRA', 'limited GPU memory', 'multi-adapter serving', 'HuggingFace', 'transformers' are all phrases users would say when needing this skill. A few natural terms are missing (e.g., 'PEFT', 'adapter training', 'quantization', 'fine-tune a model on consumer GPU'), so it falls short of the comprehensive-with-synonyms level 5.

4 / 5

Distinctiveness Conflict Risk

The LoRA/QLoRA/parameter-efficient niche is well differentiated with distinct triggers ('LoRA', 'QLoRA', '<1% of parameters'), giving minimal conflict with unrelated skills. Minor overlap risk remains with closely related skills like general LLM fine-tuning or model quantization, matching the 'mostly distinct; minor overlap risk' anchor rather than the clear-niche level 5.

4 / 5

Total

16

/

20

Passed

Validation

93%

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

CriteriaDescriptionResult

metadata_field

'metadata' should map string keys to string values

Warning

Total

15

/

16

Passed

Repository

davila7/claude-code-templates

Path

cli-tool/components/skills/ai-research/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.