huggingface-llm-trainer

This skill should be used when users want to train or fine-tune language models using TRL (Transformer Reinforcement Learning) on Hugging Face Jobs infrastructure. Covers SFT, DPO, GRPO and reward modeling training methods, plus GGUF conversion for local deployment. Includes guidance on the TRL Jobs package, UV scripts with PEP 723 format, dataset preparation and validation, hardware selection, cost estimation, Trackio monitoring, Hub authentication, and model persistence. Should be invoked for tasks involving cloud GPU training, GGUF conversion, or when users mention training on Hugging Face Jobs without local GPU setup.

Quality

88%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Critical

Do not install without reviewing

Quality

Content

77%Weight 40%Scale 1-3

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-sequenced skill body with strong validation feedback loops and real executable examples, but it is oversized and repetitive, with substantial inline detail and duplicated troubleshooting/failure-mode content that belongs in the existing reference files.

Suggestions

Consolidate the repeated Hub-push, timeout, and inline-submission warnings into a single authoritative section and have later sections reference it, cutting roughly 100-150 lines of duplication.

Move the full Common Failure Modes and inline Troubleshooting lists into references/troubleshooting.md, keeping only a one-line pointer plus the most critical gotcha (ephemeral environment / timeout) in SKILL.md.

Trim the 'What is GGUF' explanation and other concept definitions Claude already knows, leaving only the TRL/HF-Jobs-specific operational details that earn their tokens.

Dimension	Reasoning	Score
Conciseness	The ~700-line body reiterates the same warnings across many sections (ephemeral Hub push, 30-min timeout, inline/asynchronous submission recur in Key Directives, Asynchronous Guidelines, Critical Saving, Failure Modes, Troubleshooting, and Key Takeaways) and includes concept explanation ('What is GGUF'), so it is mostly efficient but could be tightened rather than fully lean.	2 / 3
Actionability	Provides fully executable, copy-paste-ready guidance: a complete hf_jobs() SFT script, exact CLI syntax with explicit wrong-vs-correct contrasts, concrete cost-estimation commands, and a real dataset-inspector URL with script_args.	3 / 3
Workflow Clarity	Multi-step flows are clearly sequenced with explicit validation checkpoints and feedback loops, notably the dataset-validation loop ('✓ READY → proceed, ✗ NEEDS MAPPING → apply code, ✗ INCOMPATIBLE → choose different'), a pre-submit verification checklist, and ordered fix sequences for OOM/timeout/hub-push failures.	3 / 3
Progressive Disclosure	References are one-level-deep, clearly signaled inline ('See references/gguf_conversion.md') and confirmed to exist, with a Resources index for navigation, but the SKILL.md itself is oversized with full hardware/timeout tables and Common Failure Modes + Troubleshooting sections that duplicate references/troubleshooting.md and references/hardware_guide.md — content that should be separate is inline.	2 / 3
	Total	10 / 12 Passed

Description

100%Weight 40%Scale 1-3

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, specific description that names concrete actions, natural trigger terms, explicit use-conditions, and a distinct niche without relying on fluff or first/second person. It is somewhat verbose in its middle enumeration but every clause names a real capability rather than padding.

Dimension	Reasoning	Score
Specificity	Lists multiple concrete capabilities ('train or fine-tune language models', 'SFT, DPO, GRPO and reward modeling', 'GGUF conversion', 'dataset preparation and validation, hardware selection, cost estimation, Trackio monitoring, Hub authentication, and model persistence') in third-person voice, matching the 'lists multiple specific concrete actions' anchor.	3 / 3
Completeness	Explicitly answers both what (training methods plus supporting tasks) and when via 'This skill should be used when users want to...' and 'Should be invoked for tasks involving...', satisfying the explicit-trigger requirement.	3 / 3
Trigger Term Quality	Includes natural terms users would actually say ('train or fine-tune language models', 'Hugging Face Jobs', 'GGUF conversion', 'cloud GPU training', 'without local GPU setup') with good coverage of common variations.	3 / 3
Distinctiveness Conflict Risk	Targets a clear niche (TRL training on Hugging Face Jobs infrastructure) with distinct triggers, making it unlikely to fire for unrelated skills.	3 / 3
	Total	12 / 12 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
skill_md_line_count	SKILL.md is long (717 lines); consider splitting into references/ and linking	Warning

	Total	15 / 16 Passed

Repository: openai/plugins
Path: plugins/hugging-face/skills/llm-trainer/SKILL.md
Commit: 11c74d6

Reviewed: 17 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.