RetentaRetenta

Evidence record · captured 2026-10-11

Hugging Face LLM Trainer

Guides fine-tuning language and vision models with TRL (SFT, DPO, GRPO, reward modeling) or Unsloth on Hugging Face Jobs cloud GPUs: dataset validation, hardware and cost estimation, UV scripts in PEP 723 format, Trackio monitoring, saving to the Hub, and GGUF conversion for local inference. Bundled example training scripts are submitted as jobs.

Official publisherUnverifiedSource reviewedBehavior not tested
Source review

Editorial assessment

This is a source-only assessment. It does not claim runtime or security testing.

What it does

Turns a fine-tuning request into a submitted Hugging Face Jobs run: a PEP 723 script with Trackio, sent inline to hf_jobs, returning a job ID, with Hub push mandatory because the container is wiped.

Best for

  • Running SFT, DPO or GRPO fine-tunes with TRL or Unsloth without a local GPU
  • Checking whether a Hub dataset matches a training method's expected columns before paying for GPU time
  • Converting a fine-tuned LoRA adapter to quantized GGUF for llama.cpp, Ollama or LM Studio

Strengths in the source

  • Has the agent run a CPU dataset inspector before GPU training; it reports READY, NEEDS MAPPING or INCOMPATIBLE per method and emits mapping code.
  • Treats persistence as a checklist: push_to_hub, a hub_model_id, and HF_TOKEN passed in secrets, since all files are deleted when the job ends.
  • Warns that the 30-minute default timeout is too short and gives per-scenario timeouts plus a 20 to 30 percent buffer.
  • Bundles estimate_cost.py and a table mapping model size to GPU flavor and hourly cost, and tells the agent to offer an estimate for jobs over an hour or $5.

Limitations

  • Spends money: HF Jobs require a paid Pro, Team or Enterprise plan and bill GPU time, and the skill tells the agent to submit the job immediately when asked to train.
  • Pushes models, checkpoints and GGUF files to Hub repos with the user's write token; repos are public unless the agent sets them private.
  • Pulls the dataset inspector and TRL example scripts from URLs at run time, and its hourly prices and library pins are snapshots that will drift.

When not to use it

  • You want training to run on your own hardware; the main flow targets HF Jobs cloud GPUs.
  • You need explicit sign-off on spend before any job starts, unless you tell the agent to stop at the cost estimate.

Reviewed by Retenta editorial on 2026-10-11; scope: source only.

Files included in this review
  • skills/huggingface-llm-trainer/SKILL.md

Evidence and boundaries

Execution · not tested

Retenta has not executed this skill.

Security · not tested

Bundles 5 scripts (scripts/train_sft_example.py, train_dpo_example.py, train_grpo_example.py, unsloth_sft_example.py, scripts/convert_to_gguf.py, scripts/estimate_cost.py, scripts/dataset_inspector.py, scripts/hf_benchmarks.py) that the agent may run. Retenta has not run them or scanned them.

Revision
c3ff942be3e07116b475253f0a8fc4f54682c1a8
License evidence
Apache-2.0Retenta preserves the license text captured with this immutable upstream snapshot; no independent legal or security review is claimed.
Last source change
2026-04-09
Independent verification
Published by Hugging Face, but not independently verified by a third-party reviewer.
Upstream repository
Open publisher repository

Compatibility

Claude Code

documented

Named in the upstream skill or repository documentation; not tested by Retenta.

OpenAI Codex

documented

Named in the upstream skill or repository documentation; not tested by Retenta.

Cursor

documented

Named in the upstream skill or repository documentation; not tested by Retenta.

Gemini CLI

documented

Named in the upstream skill or repository documentation; not tested by Retenta.

Requirements and access

  • Hugging Face account + HF_TOKEN (write)
  • Paid HF Jobs compute
  • uv / Python
  • Optional: HF MCP server for doc search

Referenced tools

No named tool is required by the reviewed SKILL.md.

Install from the source

Clones the current default branch, which may be newer than the reviewed revision c3ff942. The destination shown is Claude Code's personal skills folder; other agents read the same folder from their own skills directory.

View the SKILL.md at its source Verify the publisher's install documentation

More AI development skills →

Install the full skill folder

git clone --depth 1 https://github.com/huggingface/skills.git
cp -R skills/skills/huggingface-llm-trainer ~/.claude/skills/huggingface-llm-trainer

Thyme RAG Capsules

Make this skill’s state durable, governed, and portable.

A Retenta capsule lets agents recall prior work, append provenance-stamped progress, and carry the memory to the next run.

Create durable agent memoryUnderstand the evidence labels