hugging-face-model-trainer

Train language models using TRL (Transformer Reinforcement Learning) on fully managed Hugging Face infrastructure. No local GPU setup required—models train on cloud GPUs and results are automatically saved to the Hugging Face Hub.

Quality

62%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

High

Do not use without reviewing

Fix and improve this skill with Tessl

tessl review fix ./skills/hugging-face-model-trainer/SKILL.md

Quality

Content

63%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

This is a comprehensive and well-structured skill with strong progressive disclosure and good actionability, providing concrete code examples and clear references to supporting materials. Its main weakness is significant verbosity—critical information like timeout warnings, Hub push requirements, and approach recommendations are repeated across multiple sections, inflating token cost substantially. The workflow clarity is solid with validation checkpoints but lacks a single unified end-to-end workflow.

Suggestions

Consolidate repeated warnings (timeout, Hub push, HF_TOKEN) into a single 'Critical Configuration' section and reference it elsewhere instead of restating the same information 3-4 times.

Remove the 'When to Use This Skill' section—Claude can infer when fine-tuning is relevant from the overview and user request context.

Merge the 'Common Failure Modes' and 'Troubleshooting' sections, which currently overlap significantly, into a single concise troubleshooting reference.

Replace the GGUF 'quick conversion' placeholder with either actual inline code or just a direct reference to the conversion guide—the current hybrid approach adds tokens without adding value.

Dimension	Reasoning	Score
Conciseness	The skill is significantly verbose at ~500+ lines. It repeats information across sections (e.g., timeout warnings appear in Prerequisites, Timeout Management, Common Failure Modes, Troubleshooting, and Key Takeaways). Hub push requirements are similarly repeated 4+ times. The 'When to Use This Skill' section explains obvious use cases Claude can infer. The Key Takeaways section largely restates content already covered in detail above.	2 / 5
Actionability	The skill provides concrete, executable code examples for job submission via hf_jobs(), CLI commands, dataset validation, and training configurations. However, some sections defer to reference files (e.g., 'see references/gguf_conversion.md for complete script') rather than providing inline executable code, and the GGUF quick conversion example has a placeholder string instead of actual code. The cost estimation tool and dataset inspector are well-specified with concrete commands.	4 / 5
Workflow Clarity	The skill has clear multi-step workflows with validation checkpoints: the dataset validation workflow (inspect → check markers → apply mapping → launch), the prerequisites checklist, and the verification checklist before submission. The 'Example Workflow' for dataset validation includes an explicit feedback loop. However, there's no unified end-to-end workflow tying all steps together (prerequisites → validate → estimate cost → submit → monitor → verify results), and the post-training verification step (confirming model actually saved to Hub) is mentioned but not explicitly sequenced.	4 / 5
Progressive Disclosure	Excellent progressive disclosure structure. The SKILL.md serves as a comprehensive overview with clear, well-signaled references to 8 reference files and 6 scripts. References are consistently one level deep with descriptive labels (e.g., 'references/training_methods.md - Overview of SFT, DPO, GRPO, KTO, PPO, Reward Modeling'). The Resources section provides a clean index of all supporting materials. Content is appropriately split between the main file (workflows, quick starts, key directives) and reference files (detailed guides).	5 / 5
	Total	15 / 20 Passed

Description

61%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 clearly identifies the domain (TRL-based model training on Hugging Face infrastructure) and provides useful context about the cloud-based nature of the workflow. However, it lacks an explicit 'Use when...' clause and could be more specific about the concrete actions and training methods supported. Adding trigger terms for common fine-tuning approaches would improve discoverability.

Suggestions

Add an explicit 'Use when...' clause, e.g., 'Use when the user wants to fine-tune or train a language model using Hugging Face, TRL, or mentions cloud-based model training.'

Include specific training methods as trigger terms: 'SFT', 'RLHF', 'DPO', 'fine-tune', 'LoRA', 'reward modeling' to improve keyword coverage and specificity.

List more concrete actions such as 'configure training jobs, monitor training runs, select base models, specify datasets, and retrieve trained model artifacts'.

Dimension	Reasoning	Score
Specificity	Names the domain (training language models with TRL on Hugging Face infrastructure) and describes the general action (train models), but doesn't list specific concrete actions like fine-tuning methods (SFT, RLHF, DPO), dataset handling, or configuration options. The infrastructure details (cloud GPUs, auto-save to Hub) are more about environment than capabilities.	3 / 5
Completeness	The 'what' is reasonably clear (train language models using TRL on Hugging Face infrastructure), but there is no explicit 'when' clause. There's no 'Use when...' guidance telling Claude when to select this skill. Per rubric guidelines, missing 'Use when...' caps completeness at 3.	3 / 5
Trigger Term Quality	Includes good natural keywords like 'train language models', 'TRL', 'Transformer Reinforcement Learning', 'Hugging Face', 'GPU', and 'Hugging Face Hub'. Missing some common variations users might say like 'fine-tune', 'RLHF', 'SFT', 'DPO', 'LoRA', or 'model training'.	4 / 5
Distinctiveness Conflict Risk	Fairly distinct due to the specific mention of TRL, Hugging Face managed infrastructure, and cloud GPU training. Minor overlap risk with general Hugging Face skills or local model training skills, but the managed infrastructure and TRL specificity help differentiate it.	4 / 5
	Total	14 / 20 Passed

Validation

81%

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 — 9 / 11 Passed

Validation for skill structure

Criteria	Description	Result
skill_md_line_count	SKILL.md is long (721 lines); consider splitting into references/ and linking	Warning
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	9 / 11 Passed

Repository: administrakt0r/AI-Agents-Safe-Coding-Skills
Path: skills/hugging-face-model-trainer/SKILL.md
Commit: fbdfd63

Reviewed: 3 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.