Qwen3 Fine-tune (multi-IT)

This repository contains tools and helper code to fine-tune Qwen3 models with support for "thinking" (chain-of-thought) and mixed thinking/non-thinking training modes. It includes a compact training wrapper, configuration class, and dataset utilities to prepare and run supervised fine-tuning with LoRA and TRL/SFT workflows.

Requirements

• Python: 3.10 to 3.12
• CUDA: 21.1+ (install a matching PyTorch wheel for your CUDA runtime)

Note: adjust the PyTorch wheel URL or package versions to match your local CUDA runtime (the examples below use the cu121 wheel index).

Stack

• PyTorch: Core deep learning library used for model execution and GPU acceleration.
• Transformers: Model and tokenizer loading from Hugging Face Hub.
• PEFT: Parameter-Efficient Fine-Tuning (LoRA) utilities.
• Accelerate: Device and distributed training utilities.
• TRL (trl): Training utilities for SFT / policy learning.
• datasets: Dataset utilities and I/O.
• scikit-learn: Evaluation and metrics.
• tqdm: Progress bars.
• pandas: Data inspection and tabulation.

Set Up

Two supported ways to set up the environment: pip (virtualenv) and Poetry. Pick the one you prefer.

Important: install a matching NVIDIA driver and CUDA toolkit on your system before running Poetry so that GPU-enabled PyTorch can be installed and used.

General CUDA install steps (follow the official NVIDIA instructions for your OS and desired CUDA version):

• Install NVIDIA GPU driver (check nvidia-smi after install).
• Install the CUDA toolkit/version that matches the PyTorch build you intend to use (e.g. CUDA 12.1 for cu121).
• Verify installation:

nvidia-smi
nvcc --version

1. Using pip + virtualenv

Create and activate a virtual environment, then install packages. Adjust the PyTorch index URL to match your CUDA version if needed.

python3 -m venv .venv
source .venv/bin/activate

Install PyTorch wheels from the official PyTorch wheel index (example: CUDA 12.1 / cu121)

pip install --upgrade pip
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

Install the remaining Python dependencies

pip install numpy transformers peft accelerate scikit-learn python-dotenv jupyter trl tqdm datasets pandas

2. Using Poetry

Poetry manages dependencies and lockfiles and will install exactly what's declared in pyproject.toml.

Once system CUDA is present, simply run:

poetry install

If the resolver cannot find pre-built PyTorch wheels for your CUDA version on PyPI, configure the PyTorch wheel index locally and then run poetry install (optional):

poetry config repositories.pytorch https://download.pytorch.org/whl/cu121 --local
poetry install

Poetry will use the project pyproject.toml to install all declared dependencies and create/update poetry.lock.

References

• Mult-IT: Multiple Choice Questions on Multiple Topics in Italian:A CALAMITA Challenge