Finetune Studio

Finetune Studio

Fine-tunes an open model on your own conversations with no code. Sign in with Hugging Face, pick a model and a set of chats, choose how long to train, and the finished model lands in your account.

Made by qgallouedec 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: Finetune Studio lets you fine-tune an open language model on your own conversations without writing code. After you sign in with your Hugging Face account, you choose a model, a set of chats and a time budget, then press Train. The training job runs on Hugging Face Jobs, and the finished model is saved to your own account.

Who it suits: It suits people who have never written code and want a model that answers in the style of their own chats. The only choice is how long to train, and everything else follows a ready-made recipe.

Industry | Who uses it | How they use it | Example
Support and service | Team leads | Teach a small model the tone of their past replies | Train on a set of support chats
Writing and coaching | Coaches and creators | Make a model that answers in their own voice | Train on saved conversations
Education | Teachers | Adapt a model to a style of explanation | Train on tutoring chats
Research | ML hobbyists | Try a tested LoRA recipe across four model families | Compare two of the four models

You can skip it if you want to set learning rates, batch sizes and chat templates yourself, since the recipe fixes them.

Where to use it:

• A web browser: open the Space page on Hugging Face and sign in
• Windows, macOS or Linux: run a copy in Docker
• Hugging Face Jobs: the training run itself goes there, on your own credit

Use case | Who | What they get back
Train on your chats | Non-coders | An ordinary model saved in your Hugging Face account
Pick a model size | Hobbyists | A choice of four open models
Watch a run | Anyone | Live progress from the run's log

How to set it up: Use the online page with a Hugging Face account that can pay for jobs. A local copy needs Docker.

• A Hugging Face account, since you sign in with it
• Credit on that account for the training job
• Docker for a local copy
• A web browser

1. To use it online, open the Space page on Hugging Face and sign in.
2. To run a copy, install the Hugging Face hub package.
3. Download the Space into a local folder.
4. Move into the folder and build the Docker image.
5. Run the container with the port published.

pip install -U huggingface_hub
hf download qgallouedec/finetune-studio --repo-type space --local-dir finetune-studio
cd finetune-studio
docker build -t finetune-studio .
docker run -it -p 7860:7860 finetune-studio

How to use it: Sign in, pick one of the four models, choose a dataset of chats, choose how long you are willing to wait and click Train. The page checks the dataset size, starts the job and shows progress. You can refresh the page without losing the run.

Example: You sign in, pick Qwen3-4B, choose a set of your chats and take the Quick plan. A quick run takes about 9 to 11 minutes. When it ends, you get an ordinary model in your account and a before and after comparison.

For newcomers and experts: If you are new to this, fine-tuning means teaching an existing model your style by showing it examples. Start with the Quick plan on a small set of chats and compare the before and after.

If you know the field, the recipe is LoRA on all linear layers with a learning rate of 2e-4, assistant-only loss, packing and a 16384 token limit, with the adapter merged before pushing. Plans are token budgets, so epochs and gradient accumulation follow from a measured sample. The menu offers four models, one each from the SmolLM3, Qwen3, Phi-3.5 and gpt-oss families. The browser reads from the Hub directly, and only starting a run needs the Space's own server.

Sign in with Hugging Face
Pick a model, chats and a time budget
The Space checks tokens and builds the recipe
The job runs on Hugging Face Jobs
The merged model lands in your account

Good to know: Nothing is stored on the Space's side. Run details live in the job's labels, progress in its log and results in your model repository. The notes in the repository say a learning rate of 5e-5 transferred nothing, while 1e-4 and above worked.

Title: Mean tokens per conversation across real datasets
Shortest average: 161 tokens
Longest average: 1146 tokens

• License: apache-2
• Cost to run: a Hugging Face account with credit for Jobs, as the training runs on your own credit
• Data it reaches: your Hugging Face account, the datasets and chats you pick, and the model repository it creates
• Not for: hand-tuning learning rates, batch sizes or chat templates, since one curated recipe is used

On Hugging Face, 13 November 2026

• Likes: 6
• Built with: docker
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/qgallouedec/finetune-studio
• Files: https://huggingface.co/spaces/qgallouedec/finetune-studio/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.