ArmanXAI/bert-base-agnews-news-classifier

BERT-base AG News News Classifier

Fine-tuned bert-base-uncased on AG News for 4-class news topic classification. 94.74% accuracy · 94.75 macro F1 — trained on RTX 4060 with fp16 in under an hour.

---

Model Description

This model is a fine-tuned version of bert-base-uncased on the AG News dataset, a benchmark corpus of 120,000 news headlines across four topic categories.

It classifies a short news headline or sentence into one of four classes:

ID	Label
0	World
1	Sports
2	Business
3	Sci/Tech

Built as part of the DevelopersHub AI/ML Engineering Internship (Advanced Track, 2026).

---

Intended Use

• Primary: Single-sentence or short headline news topic classification (English).
• Good for: News aggregators, content tagging pipelines, NLP demos.
• Out of scope: Long-form articles (>512 tokens), non-English text, opinion classification.

---

Training Details

Dataset

Split	Size
Train	120,000
Test	7,600

Classes are perfectly balanced (30,000 / 1,900 per class).

Hyperparameters

Parameter	Value
Base model	bert-base-uncased
Learning rate	2e-5
Batch size	16
Epochs	3
Mixed precision	fp16
Optimizer	AdamW
Warmup ratio	0.1
Max sequence len	128

Hardware

Trained on NVIDIA RTX 4060 Laptop GPU (8 GB VRAM) using HuggingFace Trainer.

---

Evaluation Results

Overall

Metric	Value
Accuracy	94.74%
Macro F1	0.9475
Weighted F1	0.9474

Per-Class F1

Class	F1
World	0.95
Sports	0.99
Business	0.92
Sci/Tech	0.93

Baseline Comparison

Model	Accuracy
TF-IDF + Logistic Reg.	~91%
This model (BERT)	94.74%

---

How to Use

Quick inference (HuggingFace pipeline)

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="ArmanXAI/bert-base-agnews-news-classifier"
)

result = classifier("NASA launches new telescope to study deep space phenomena.")
print(result)
# [{'label': 'Sci/Tech', 'score': 0.984}]

Manual inference

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "ArmanXAI/bert-base-agnews-news-classifier"
tokenizer  = AutoTokenizer.from_pretrained(model_name)
model      = AutoModelForSequenceClassification.from_pretrained(model_name)

labels = ["World", "Sports", "Business", "Sci/Tech"]

def classify(text: str) -> dict:
    inputs = tokenizer(text, return_tensors="pt",
                       truncation=True, max_length=128)
    with torch.no_grad():
        logits = model(**inputs).logits
    pred  = torch.argmax(logits, dim=-1).item()
    score = torch.softmax(logits, dim=-1)[0][pred].item()
    return {"label": labels[pred], "score": round(score, 4)}

print(classify("Apple unveils M4 chip with major AI performance gains."))
# {'label': 'Sci/Tech', 'score': 0.9821}

---

Limitations

• English only — performance drops significantly on other languages.
• Short text — trained on headline-length inputs (~7–15 tokens); degrades on paragraphs > 256 tokens.
• Temporal drift — AG News corpus is from 2004–2005; accuracy may drop on very recent entity names, products, or events.
• 4 coarse classes — not suitable for fine-grained topic classification.

---

Training Code

See the training notebook in the notebooks/ folder of this repository, or on GitHub: DevelopersHub-AI-Internship-Advance-Tasks

---

About the Author

Arman Adil Mangat — AI/ML Engineer, BS AI '26, UMT Lahore 🇵🇰

Specializing in Embedded AI + NLP. Building GENIE — a fully offline voice assistant on ESP32-S3 (Whisper + Mistral 7B, no cloud, $22 hardware).

• 🐙 GitHub : ArmanAdilMangat
• 🤗 HF : ArmanXAI
• 💼 LinkedIn: arman-adil-mangat

---

Citation

@misc{mangat2026bertagNews,
  author    = {Arman Adil Mangat},
  title     = {BERT-base Fine-tuned on AG News for News Topic Classification},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/ArmanXAI/bert-base-agnews-news-classifier}
}

---

License

MIT — free to use, modify, and distribute with attribution.