Training Hub plugin
A plugin of four skills that guides setup, GPU memory estimation and fine-tuning runs of LLMs through the Training Hub interface.
- 105 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- The skills launch shell scripts and long training jobs that load GPUs and write models and logs to disk.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/training-hubDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The skills launch shell scripts and long training jobs that load GPUs and write models and logs to disk.
This is third-party code. Review the repository files before installing.
What it does
Training Hub is a Red Hat AI Innovation Team library that unites language model fine-tuning algorithms behind one interface: SFT, continual learning OSFT, LoRA, GRPO, GEPA and embedding fine-tuning. The repository carries a plugin for Claude Code and Codex with skills, and this card describes that plugin. The `setup-guide` skill detects the environment and helps set up training from scratch, `memory-estimation` estimates GPU memory before a run to see whether a model fits, `training-guide` launches a job from a saved configuration and `training-hub-guide` suggests an algorithm, hyperparameters and how to read out-of-memory errors and loss curves. The `th_detect.sh`, `th_estimate.sh` and `th_train.sh` scripts carry out the skills' steps.
Who it is for. Machine learning engineers who fine-tune language models and want an agent to pick an algorithm, estimate memory and launch training.
Good fit when
- Before a run you need to know whether a model fits into the memory of the available GPUs.
- Training is set up for the first time and the agent detects the environment and installs the right extras.
- You choose between SFT, OSFT and LoRA for the data size and hardware.
- Training crashes with an out-of-memory error and the causes need sorting out.
Not a fit when
- There is no GPU: the main algorithms target CUDA and only GEPA and embedding fine-tuning run on CPU.
- You need full library documentation: the site marks it as work in progress.
- Training runs through another library: the plugin knows Training Hub only.
Example request
Estimate whether the model fits on 2 GPUs of 80 GB for LoRA fine-tuning and suggest a configuration.Limitations
The plugin helps work with the library, which is installed separately with `pip install training-hub` and extras per algorithm: `lora`, `grpo`, `gepa`, `embedding`. SFT and GRPO list multiple GPUs and nodes, LoRA on Unsloth can use a single GPU and GEPA runs on CPU through an API. The skills launch the shell scripts `th_detect.sh`, `th_estimate.sh` and `th_train.sh`, and training takes a lot of time and memory. The library documentation is marked as work in progress.
How to disable. Remove the plugin from Claude Code or Codex and the library with `pip uninstall training-hub`. Saved configurations and training results are removed by hand.
FAQ
How do I install the library?
Run `pip install training-hub`. Add the `[lora]` extra for LoRA, `[grpo]` for GRPO, `[gepa]` for GEPA and `[embedding]` for embeddings.
Which algorithms are supported?
SFT, continual learning OSFT, LoRA with SFT, LoRA with GRPO, full GRPO, GEPA and embedding fine-tuning.
What does the memory-estimation skill do?
Estimates GPU VRAM use for a configuration before a run and helps tell whether a model fits and how to allocate cards.
When should I call setup-guide?
On the first training setup or when `training_hub` is not yet installed in the current environment. The skill detects the environment with the `th_detect.sh` script.
Which agents can use it?
The repository has plugin manifests for Claude Code and Codex, and skills sit in `.claude/skills`.
Discussion
No questions or reviews yet.
Related
PluginsA CLI for fine-tuning language models from one YAML: LoRA, QLoRA, DPO and GRPO, layer streaming for 8B on a 4 GB GPU, export to GGUF
A synthetic data Python framework with a Claude Code and Codex plugin: blocks chain into YAML flows that generate datasets
Ledidi designs minimal edits to DNA, RNA and protein sequences for a target model output; its skill guides agents.
Probabl skills for machine learning experiments: a scikit-learn and skrub pipeline, evaluation, audit and a link to Skore Hub
Official Hugging Face skills: Hub operations via the hf CLI, datasets, model training, Spaces, evals and deployment
Two skills from Stas Bekman's book: training large models on a cluster and an acceptance run of a GPU cluster with a report.
README in short
The repository contains the Training Hub library with examples and documentation, plugin manifests for Claude Code and Codex and four skills in `.claude/skills/`. The license is Apache-2.0.