24-679-distilbert-player-position

DistilBERT Premier League Position Classifier

Purpose

This model predicts a Premier League player's position from a text description.

Classes:

• 0 = Goalkeeper
• 1 = Defender
• 2 = Midfielder
• 3 = Forward

Dataset

Dataset: kadireks/2026-24679-text-dataset

• Train: 1050 examples
• Validation: 15 examples
• Test: 15 examples

The text column is used as input and label is the classification target.

Preprocessing

AutoGluon MultiModal handles DistilBERT tokenization and batching. Inputs are limited to 128 subword tokens. No additional random text augmentation was applied during fine-tuning.

Training

• Model: DistilBERT (distilbert-base-uncased)
• Task: multiclass classification
• Fine-tuning: full model
• Learning rate: 2e-5
• Weight decay: 0.01
• Maximum epochs: 5
• Batch size: 8
• Seed: 24679
• Best checkpoint selected using validation accuracy

Results

• Validation accuracy: 0.5333
• Test accuracy: 0.7333
• Weighted F1: 0.7222
• Misclassified test examples: 4 / 15

A confusion matrix and per-class metrics are included in the notebook.

Error Analysis

The model misclassified four of the fifteen test examples. Two defenders were predicted as midfielders, likely because their descriptions emphasized passing, ball control, and build-up play. One midfielder was predicted as a forward because its description focused on progressive attacking play. One goalkeeper was predicted as a forward despite containing some goalkeeper-specific language.

Limitations

The validation and test sets each contain only 15 examples, so the reported performance has substantial uncertainty. Descriptions of different positions can also contain overlapping language, especially for defenders and midfielders. The model may not generalize to other leagues or different styles of player descriptions.

Ethical Considerations

This model predicts playing position only. It should not be used to judge player quality, value, employment decisions, or personal characteristics.

Hardware and Compute

Training was completed in Google Colab using one GPU with approximately 14.56 GB of GPU memory and 2 CPU cores. The run used AutoGluon 1.6.1, PyTorch 2.11.0, and was limited to five epochs with seed 24679.

License

MIT

Acknowledgments

The workflow was adapted from the DistilBERT fine-tuning notebook provided for CMU 24-679.

AI Usage Disclosure

ChatGPT was used as a coding copilot to assist with syntax, library APIs, and adapting the course-provided DistilBERT fine-tuning code to this notebook. I used it to help identify the changes needed for the selected dataset and workflow. ChatGPT also helped draft and polish the Model Card based on my inputs, implementation choices, and final results.