ZeroGPU Fast Mini-Trainer

ZeroGPU Fast Mini-Trainer

Fine-tunes small SmolLM models with a LoRA adapter on a ZeroGPU Space, using your own dataset. Each person signs in and spends their own daily GPU quota. For developers new to fine-tuning.

Made by Alibrown on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: This is a small Gradio web page that teaches a tiny SmolLM model something new in a few minutes. You pick a base model, name a dataset and a target repo, then press a button. The result is a LoRA adapter, a small add-on for the model, instead of a rebuilt model.

Each person signs in with Hugging Face and uses their own daily ZeroGPU quota. The quota belongs to the signed-in user and not to the Space.

Who it suits: It suits people who want a first hands-on run of fine-tuning on a tiny model, with the terms explained in plain words.

Industry | Who uses it | How they use it | Example
Machine learning | Fine-tuning beginners | Train a first adapter on a small SmolLM2 model | A learner trains on the small test dataset from the repository
Software | Developers | Teach a small model an answer style from question and answer pairs | A developer prepares a train.jsonl of examples and runs it
Education | Teachers | Show what epochs, loss and LoRA rank do | A teacher runs two settings and compares the loss values
Research | Hobbyists | Try quick LoRA experiments within a daily GPU limit | A hobbyist tests a dataset before a longer run elsewhere

You can skip it if you need to train large models or you have no Hugging Face account.

Where to use it: You use it in these places.

• A web browser, on the Space's page on Hugging Face, signed in with your account
• Your own Space created from these files, with ZeroGPU hardware
• A local copy on Windows, macOS or Linux, though the source describes only the ZeroGPU setup

Use case | Who | What they get back
Train a first adapter | Beginners | A LoRA adapter for a SmolLM2 model, with a link at the end
Test a dataset | Developers | Loss values in the status box that show whether the model is learning
Teach the terms | Teachers and students | A hands-on run that shows base model, dataset, adapter and epoch

How to set it up: You need a Hugging Face account, since the page signs you in and uses your ZeroGPU time.

• A Hugging Face account
• A dataset with your examples, private or public
• Optionally, an empty model repo for the result
• Python packages from requirements.txt, for a local copy

1. Create a new Space with the Gradio SDK and ZeroGPU hardware.
2. Upload app.py, requirements.txt and README.md, and wait until the Space shows Running. Keep README.md, because it holds the Space's settings.
3. Create a dataset, private is fine, and upload train.jsonl.
4. Optionally create an empty private model repo to receive the result.
5. For a local copy instead, run the commands below.

pip install -U huggingface_hub
hf download Alibrown/ZeroGPU-Mini-Trainer --repo-type space --local-dir ZeroGPU-Mini-Trainer
cd ZeroGPU-Mini-Trainer
pip install gradio==6.31.0
pip install -r requirements.txt
python app.py

How to use it: Open the Space and click Sign in with Hugging Face. Choose a base model, and start with the 135M one. Enter a dataset as a full URL or as name/dataset, then a target repo or leave it empty. Press the training button, labeled Training starten in the guide. The status box shows what happens, and at the end it shows the link to the finished adapter.

The dataset is a JSONL file with one example per line. The formats are detected automatically.

• messages: the recommended format, a list of user and assistant turns
• text: one whole example text per line
• instruction and output: also prompt and response, or question and answer

{"messages": [{"role": "user", "content": "Hi"}, {"role": "assistant", "content": "Hallo!"}]}

Setting | What it does
Max GPU seconds | Sets how long training may run. A free account has about 5 minutes a day, and 150 seconds leaves room for a second try.
Max examples | Sets how many rows of the dataset are used
Epochs | One pass over all examples. The guide suggests 2 or 3 for small datasets and 1 for large ones.
LoRA rank | Sets the size of the add-on. The guide calls 16 a good start.
Learning rate | The guide says 2e-4 fits LoRA almost always

To use the trained adapter, load it on top of the base model.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "HuggingFaceTB/SmolLM2-135M-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base)
model = PeftModel.from_pretrained(model, "dein-name/dein-modell-repo")

Example: You choose SmolLM2-135M-Instruct, enter your dataset name, leave the target repo empty and press the training button with 150 GPU seconds. The status box reports progress and, at the end, a link to your finished adapter.

For newcomers and experts: Fine-tuning teaches a model that already works some new behavior from examples. A LoRA adapter does this by training only a small add-on instead of the whole model, which is quick and cheap. Try the small test dataset and the 135M model first, and watch the loss value fall in the status box.

Under the hood, the guide describes a Gradio app that runs LoRA training on SmolLM2 models, with ZeroGPU time billed to the signed-in user. The adapter can be merged into a normal model afterward with merge_and_unload. The guide suggests rank 16 and learning rate 2e-4 as defaults.

Sign in with Hugging Face
Pick a base model and a dataset
LoRA training runs on ZeroGPU
The adapter goes to your target repo
You load it on top of the base model

Good to know: The daily GPU quota is the main limit, and a few usual errors have simple fixes.

• Quota error: the day's time is used up, so wait a day or set fewer seconds.
• Columns not recognized: change the dataset format or name the column in the text column field.
• Access denied: allow read and write rights at sign-in.
• Training changes little: use more and cleaner examples.
• License: agpl-3.
• Cost to run: none named beyond the ZeroGPU quota of the signed-in Hugging Face account.
• Data it reaches: your Hugging Face account, your dataset and the target repo where the adapter is stored.
• Not for: rebuilding a whole model, since it trains only a small LoRA add-on for small SmolLM models.

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: gradio
• License: agpl-3.0

Links

• Space: https://huggingface.co/spaces/Alibrown/ZeroGPU-Mini-Trainer
• Files: https://huggingface.co/spaces/Alibrown/ZeroGPU-Mini-Trainer/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.