CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

huggingface-llm-trainer

Train or fine-tune language and vision models using TRL (Transformer Reinforcement Learning) or Unsloth with 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, model selection/leaderboards and model persistence. Use for tasks involving cloud GPU training, GGUF conversion, or when users mention training on Hugging Face Jobs without local GPU setup.

Quality

78%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Critical

Do not install without reviewing

Fix and improve this skill with Tessl

tessl review fix ./skills/huggingface-llm-trainer/SKILL.md
SKILL.md
Quality
Evals
Security

Quality

Content

63%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.

Operationally rich and highly actionable, with validation checkpoints and copy-paste-ready examples that clearly encode non-obvious Hugging Face Jobs constraints. Its main weaknesses are length and redundancy — duplicated timeout/Hub-push warnings, overlapping failure-modes/troubleshooting sections, inline content that mirrors dedicated reference files, an unreferenced reference file, and a 'Three Approaches' heading that lists four.

Suggestions

Merge 'Common Failure Modes' and 'Troubleshooting' into a single section (or move the bulk into references/troubleshooting.md and keep only a one-line-per-issue pointer) — they currently cover the same OOM, timeout, dataset-format, and Hub-push failures twice.

State the timeout, Hub-push, and Trackio requirements once each (e.g., in a single 'Critical Settings' section) and reference them from elsewhere, instead of repeating full explanations in Prerequisites, Timeout Management, Key Takeaways, and the failure sections.

Fix the section that says 'Quick Start: Three Approaches' but lists four, correct the 'Load these scripts for correctly:' and 'Example -- choosing' typos, and either link references/reliability_principles.md from the Resources list or remove it from the bundle.

DimensionReasoningScore

Conciseness

The body is ~740 lines with real redundancy: timeout warnings appear in "Critical Settings", "Timeout Management", "Common Failure Modes", and "Key Takeaways"; Hub-push instructions repeat in at least four sections; and "Common Failure Modes" vs "Troubleshooting" are near-duplicates of each other. Most content is genuinely non-obvious operational knowledge (ephemeral environment, CLI flag ordering, $HF_TOKEN secrets syntax), so it is not anchor-2 padding, but the duplication and stray filler ("This is useful for managing multiple jobs with the same configuration…", "The script is fast, and will usually complete synchronously.") fit anchor 3: mostly efficient but could be tightened.

3 / 5

Actionability

The content is highly executable — complete inline `hf_jobs("uv", {...})` payloads with PEP 723 headers, correct-vs-wrong CLI syntax blocks, DPO mapping code, and exact inspector output markers (✓ READY / ✗ NEEDS MAPPING / ✗ INCOMPATIBLE). It falls short of anchor 5 because some snippets are non-executable placeholders ("script": "<see references/gguf_conversion.md for complete script>") and a few pointers end in "etc."; these are minor gaps, matching anchor 4.

4 / 5

Workflow Clarity

The end-to-end sequence is explicit and well-ordered: prerequisites checklist (auth, dataset, critical settings) → dataset validation with a cheap CPU pre-check → script creation from templates → immediate submission → reporting job ID/monitoring URL → wait for user. Validation checkpoints exist (dataset inspector, pre-submission verification checklist, cost estimator with timeout buffer), and error-recovery loops are documented in "Common Failure Modes". It stays at 4 rather than 5 because the structure has navigational defects: "Quick Start: Three Approaches" actually presents four approaches, "Common Failure Modes" and "Troubleshooting" duplicate each other so it is unclear which governs, and small errors ("Load these scripts for correctly:", "Example -- choosing") mar the guidance.

4 / 5

Progressive Disclosure

References are one level deep, all cited paths resolve to real files, and a Resources section catalogs them — but the ~740-line body inlines substantial content that duplicates dedicated bundle files (inline "Troubleshooting" alongside references/troubleshooting.md, inline GGUF conversion walkthrough alongside references/gguf_conversion.md, Trackio defaults repeated across sections), and references/reliability_principles.md exists in the bundle yet is never referenced from SKILL.md. This matches anchor 3 (content that should be separate is inline) more than anchor 4's 'most content appropriately placed'.

3 / 5

Total

14

/

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: concrete capabilities, explicit third-person voice, and a well-formed 'Use when…' clause with multiple natural trigger phrases. Its weaknesses are the feature-inventory phrasing of the middle sentence and a few missing natural synonyms (LoRA, RLHF, instruction tuning) that would round out trigger coverage.

DimensionReasoningScore

Specificity

Quotes like "Train or fine-tune language and vision models using TRL…", "SFT, DPO, GRPO and reward modeling", and "GGUF conversion for local deployment" list several concrete actions with comprehensive coverage, but the middle sentence ("Includes guidance on the TRL Jobs package, UV scripts with PEP 723 format, dataset preparation and validation…") reads as a topical feature inventory rather than specific actions. This sits between anchor 4 (several specific actions, minor gaps) and anchor 5; the inventory-style phrasing keeps it at 4.

4 / 5

Completeness

Both questions are answered explicitly: the 'what' is concrete ("Train or fine-tune… using TRL or Unsloth… Covers SFT, DPO, GRPO and reward modeling… plus GGUF conversion") and the 'when' gives concrete trigger phrases ("Use for tasks involving cloud GPU training, GGUF conversion, or when users mention training on Hugging Face Jobs without local GPU setup"). This matches anchor 5 exactly; anchor 4 would require the 'when' to be less explicit than it is.

5 / 5

Trigger Term Quality

Natural phrases are present — "train or fine-tune", "cloud GPU training", "GGUF conversion", "training on Hugging Face Jobs without local GPU setup" — matching how users would phrase requests. A few common natural terms are missing (e.g., "LoRA", "RLHF", "instruction tuning", "quantize"), which fits anchor 4 (good coverage, a few natural terms missing) rather than anchor 5's comprehensive synonym coverage.

4 / 5

Distinctiveness Conflict Risk

The description carves a clear niche via named tooling (TRL, Unsloth, Hugging Face Jobs, GGUF), but the breadth creates minor overlap risk with plausible sibling skills (a dedicated GGUF-conversion skill, a general Hugging Face Hub skill, or a standalone Unsloth skill would all match some triggers). Fits anchor 4 (mostly distinct, minor overlap with closely related skills) rather than anchor 5's minimal conflict risk.

4 / 5

Total

17

/

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

skill_md_line_count

SKILL.md is long (739 lines); consider splitting into references/ and linking

Warning

Total

15

/

16

Passed

Repository

huggingface/skills

Path

skills/huggingface-llm-trainer/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.