Quick Start

Quick Start

Requirements

• Python 3.12 recommended; create an isolated environment.
• Internet access on first run to download distilbert-base-uncased and Python packages.
• CPU is possible but slower; CUDA needs a compatible PyTorch build and driver.

Windows PowerShell

py -3.12 -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
python -m pip install -r requirements_distilbert.txt
python DistilBERT_Colab_Fixed_2026.py

requirements_distilbert.txt is a streamlined optional install list. The original requirements.txt additionally includes TRL/PEFT/bitsandbytes for the unfinished Phi-2 example.

Jupyter / Colab

Open DistilBERT_Colab_Fixed_2026.ipynb and run the cells in order. In a notebook, %pip install -r requirements_distilbert.txt requires that requirements file be uploaded into the notebook working directory.

Expected files after execution

• support_classifier/ training checkpoints and support_classifier/best_model/ (saved classifier + tokenizer)
• support_classifier_frozen/ for frozen-backbone experiment logs

These are local outputs, not committed sample models. Training results and accuracy depend on environment and training run. See Artifacts-and-Checkpoints and Troubleshooting.