DistilBERT Football Sentiment — Positive vs Negative

Purpose

Fine-tune a compact transformer (DistilBERT) to classify short football-related comments as positive (1) or negative (0). This supports a course assignment on text modeling and evaluation.

Dataset

Source: james-kramer/football_news on Hugging Face.
Schema: text (string), label (0/1).
Task: Binary sentiment classification (0=negative, 1=positive).
Splits: Stratified 80/10/10 (train/val/test) created in this notebook.
Cleaning: Strip text, drop empty/NA rows.

Preprocessing

Tokenizer: distilbert-base-uncased (uncased), max_length=256, truncation.
Label mapping: {0: "negative", 1: "positive"}.

Training Setup

Base model: distilbert-base-uncased
Epochs: 5
Batch size: 16
Learning rate: 3e-05
Weight decay: 0.01
Warmup ratio: 0.1
Early stopping: patience = 2 (monitor F1 on validation)
Seed: 42
Hardware: Google Colab (GPU)

Metrics (Held-out Test)

{
  "eval_loss": 0.0029852271545678377,
  "eval_accuracy": 1.0,
  "eval_precision": 1.0,
  "eval_recall": 1.0,
  "eval_f1": 1.0,
  "eval_runtime": 0.3123,
  "eval_samples_per_second": 352.273,
  "eval_steps_per_second": 22.417,
  "epoch": 4.0
}

Confusion Matrix & Errors

The Colab notebook includes a confusion matrix for validation and test, plus a short error analysis with example misclassifications and hypotheses (e.g., injury news phrased neutrally but labeled negative).

	Pred 0	Pred 1
True 0	55	0
True 1	0	55

Brief Error Analysis (Concrete Examples & Hypotheses)

No misclassifications were observed in the held-out test split (confusion matrix = perfect).
However, given the very small dataset size (~30 examples), this likely reflects overfitting rather than true robustness.

Limitations & Ethics

Dataset size and labeling style can lead to unstable metrics; neutral/ambiguous tone is hard.
Sports injury and team-management news may bias wording and labels.
For coursework only; not for production or sensitive decisions.

Reproducibility

Python: 3.12
Transformers: >=4.41
Datasets: >=2.19
Seed: 42

License

Code & weights: MIT (adjust per course guidelines)
Dataset: see the original dataset's license/terms

AI Assistance Disclosure

GenAI tools assisted with notebook structure and documentation; modeling choices and evaluation were implemented and verified by the author.

Downloads last month: 1

Safetensors

Model size

67M params

Tensor type

F32