Fine Tuning Example

In this tutorial, we demonstrate how to perform post-training / fine-tuning using Nebius Token Factory

Fine Tuning Example
1. Get the code
2. Install dependencies (If running locally)
3 - Create .env file
4 - Running the code
5 - Fine tuning
6 - Your fine-tuned model in Nebius Token Factory
7 - Using your fine-tuned model in Playground
8 - Use the fine-tuned model using an API
References
Dev Notes

1. Get the code

- run without any local setup!

git   clone    https://github.com/nebius/token-factory-cookbook/
cd    post-training/fine-tuning-1

2. Install dependencies (If running locally)

If using uv (preferred)

uv sync
uv add --dev ipykernel   # only when setting up UV for first time
uv run python -m ipykernel install --user --name="fine-tuning-1" --display-name "fine-tuning-1"
# select this kernel when running in jupyter / vscode

If using Conda

conda  create  -n fine-tuning-1  python=3.12
conda activate  fine-tuning-1
pip install -r requirements.txt
python -m ipykernel install --user --name="fine-tuning-1" --display-name "fine-tuning-1"

If using pip

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python -m ipykernel install --user --name="fine-tuning-1" --display-name "fine-tuning-1"
# select this kernel when running in jupyter / vscode

3 - Create .env file

Create a .env file in the project root and add your Nebius API key:

cp env.example .env

NEBIUS_API_KEY=your_api_key_here

4 - Running the code

Using VSCode

open the notebook: fine_tune_llama.ipynb
run with custom jupyter kernel we created (fine-tuning-1)

Using uv

uv run --with jupyter jupyter lab fine_tune_llama.ipynb

Using standard python/pip

jupyter lab

5 - Fine tuning

You can see fine tuning jobs' status on the post training dashboard

The fine tuned model will be saved locally into models-checkpoints directory.

6 - Your fine-tuned model in Nebius Token Factory

Find your fine-tuned models in models --> private section

7 - Using your fine-tuned model in Playground

Try your new shiny model in the playground!

8 - Use the fine-tuned model using an API

See our API examples

And documentation

References

Dev Notes

How to setup a uv project

uv init --python=3.12
uv add openai datasets python-dotenv  pandas  seaborn  jinja2
uv add --dev ipykernel

# create a requirements.txt file
uv export --frozen --no-hashes --no-emit-project --no-default-groups --output-file=requirements.txt

Notebooks in this recipe:

fine_tune_llama.ipynb · nbviewer ↗