Parameter	Value
Epochs	3
Batch Size	16
Learning Rate	2e-5
Max Sequence Length	128
Warmup Steps	500
Optimizer	AdamW
Weight Decay	0.01
LR Scheduler	Linear Warmup + Linear Decay
Gradient Clipping	Max Norm = 1.0

Dataset Information

Split	Samples
Train	120,000
Test	7,600

Dataset source:

:contentReference[oaicite:3]{index=3}

Performance

Classification Report

Class	Precision	Recall	F1-Score
World	0.97	0.95	0.96
Sports	0.98	0.99	0.99
Business	0.92	0.91	0.92
Sci/Tech	0.91	0.93	0.92

Usage

Install

pip install transformers torch

Inference

python

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="theshekslaw/ag_news_bert_full"
)

result = classifier(
    "Apple announces new MacBook Pro with M3 chip"
)

print(result)

Example output:

[{'label': 'Sci/Tech', 'score': 0.9895}]

Example Predictions

Text	Prediction
Apple announces new MacBook Pro with M3 chip	Sci/Tech
Federal Reserve raises interest rates	Business
UN calls for ceasefire in ongoing conflict	World
Real Madrid wins Champions League final	Sports

Training Approach

The model was fine-tuned using a custom PyTorch training loop instead of the Hugging Face Trainer API.

Key components:

Dynamic padding with DataCollatorWithPadding
Linear learning-rate scheduling
AdamW optimizer
Mixed precision training (FP16)
GPU training on NVIDIA T4

Limitations

The model is trained only on AG News categories.
It may struggle with:
- ambiguous headlines
- multi-topic articles
- very short text
- domain-specific financial or scientific jargon

Citation

bibtex

@misc{ag_news_bert_classifier,
  title={AG News BERT Classifier},
  author={Abhishek Pandey},
  year={2026},
  publisher={Hugging Face}
}