Tessl • The package manager for agent skills

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.

Quality

82%

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

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-structured, highly actionable reference: executable code with concrete configuration values, useful selection tables, and clearly signaled one-level-deep reference files. Weaknesses are minor — a few code examples with undefined imports/variables, evaluation appearing only as an implicit best practice rather than a workflow checkpoint, and advanced method/integration/benchmark detail inlined in SKILL.md that would better live in the advanced-usage reference.

Suggestions

Move the IA3, Prefix Tuning, Axolotl, and vLLM integration sections plus the performance benchmark tables into references/advanced-usage.md, keeping SKILL.md focused on the core LoRA/QLoRA workflow and linking out — this improves both conciseness and progressive disclosure.

Fix executability gaps in the code examples: import torch before the data_collator lambda in the quick start, and define or note the prerequisites (model, dataset, prompts) in the SFTTrainer and vLLM snippets.

Promote evaluation into an explicit workflow checkpoint (e.g., a numbered step "Evaluate on held-out data with eval results before merging" in the loading/merging section) instead of a best-practices bullet.

Dimension	Reasoning	Score
Conciseness	Largely efficient — dense tables, commented code, no explanations of basics Claude already knows — but the three installation variants, the MMLU quality/benchmark tables, and per-method code duplicating the methods comparison table are trimmable padding.	4 / 5
Actionability	Mostly executable, copy-paste-ready guidance (full LoRA/QLoRA configs, rank/alpha selection tables, per-architecture target_modules), with minor gaps: the quick-start data_collator uses torch.stack without importing torch, and the SFTTrainer and vLLM examples reference undefined variables (model, dataset, prompts).	4 / 5
Workflow Clarity	The lifecycle sequence (install → configure → train → save/merge → serve) is logically ordered with an error-recovery section and a print_trainable_parameters() checkpoint, but validation is only implicit — "Evaluate on held-out data before merging" appears as a best-practice bullet rather than an explicit workflow checkpoint.	4 / 5
Progressive Disclosure	Both referenced files (references/advanced-usage.md, references/troubleshooting.md) are real, clearly signaled with descriptions, and one level deep with no nested references — but substantial advanced material (IA3, Prefix Tuning, Axolotl/vLLM integrations, benchmark tables) is inlined in the 421-line body where it belongs in advanced-usage.md.	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: it states concrete capabilities in third person and includes an explicit, multi-condition "Use when" clause with specific triggers (model sizes, memory limits, parameter budgets). Keyword coverage is good though not exhaustive of natural synonyms, and it is clearly distinguishable from adjacent training skills with only minor overlap risk.

Dimension	Reasoning	Score
Specificity	Names the domain and several concrete capabilities ("fine-tuning for LLMs using LoRA, QLoRA, and 25+ methods", "train <1% of parameters", "multi-adapter serving") in third-person voice, but coverage is not comprehensive — adapter merging, evaluation, and serving details are absent.	4 / 5
Completeness	Explicitly answers both: 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... or for multi-adapter serving") with concrete trigger phrases.	5 / 5
Trigger Term Quality	Good natural keyword coverage ("fine-tuning", "LoRA", "QLoRA", "GPU memory", "7B-70B", "multi-adapter serving"), but misses common variations users would say such as "fine-tune a model", "PEFT", or "adapters" on their own.	4 / 5
Distinctiveness Conflict Risk	Clear PEFT/LoRA niche with specific model-size and memory-constraint triggers, but minor overlap risk with a general fine-tuning or transformers-training skill that covers the same models.	4 / 5
	Total	17 / 20 Passed

Validation

87%

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 — 14 / 16 Passed

Validation for skill structure

Criteria	Description	Result
metadata_version	'metadata.version' is missing	Warning
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	14 / 16 Passed

Repository: synthetic-sciences/openscience
Path: backend/cli/skills/ml-training/peft/SKILL.md
Revision: 6a85742

Reviewed: 22 days ago

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.

peft-fine-tuning

Table of Contents

Is this your skill?