joseeangel/bert-base-uncased-agnews-topic

bert-base-uncased adapted to Topic classification (AG News) with Full fine-tuning.

Produced for the assignment U2T01 - Adapting BERT for NLP tasks (Trends in Data Science). The delivered method was chosen by measurement, not by default: the table below is the full set of adaptation methods trained for this task, and this repository holds the winner.

Intended use

Assigning one of four news topics (World, Sports, Business, Sci/Tech) to a short English news headline plus lead sentence.

Out of scope: any use where an error carries real cost without a human in the loop, and any language or domain other than the one above.

Training data

• Dataset: fancyzhx/ag_news
• License: Custom (AG News corpus, academic use)
• Citation: Zhang, Zhao & LeCun (2015), Character-level Convolutional Networks for Text Classification
• Splits used: {"train": 20000, "dev": 5000, "test": 7600}

Adaptation method

Method	Full fine-tuning
Trainable parameters	109,485,316 of 109,485,316 (100.0%)
Head learning rate	0.001
Encoder learning rate	2e-05
Epochs / batch size	2 / 32
Max sequence length	128
Scheduler	linear with 10% warmup
Seed	42
Hardware	Apple M5 (mps)
Wall-clock training time	16.6 min

A freshly initialised head and pretrained encoder weights are trained in two parameter groups with separate learning rates; a single shared rate either starves the head or destroys pretrained features.

Evaluation

Held-out test split, never seen during training or model selection.

Metric	Value
Accuracy	93.01
Macro F1	93.00

What every method scored on this task

Method	Trainable params	Share of model	Accuracy	Macro F1	Train time (min)
Feature-based (logistic regression, mean-pooled)	3,076	0.0%	90.18	90.18	2.9
Feature-based (linear probe, pooler as head)	593,668	0.542%	89.61	89.59	13.6
Feature-based (linear SVM)	3,076	0.0%	89.21	89.20	0.1
Feature-based (logistic regression)	3,076	0.0%	88.61	88.59	2.0
Feature-based (random forest)	0	0.0%	84.87	84.79	0.3
Feature-based (MLP probe)	395,780	0.36%	84.58	84.57	14.8
Feature-based (linear probe)	3,076	0.003%	82.09	82.04	12.5
Partial FT (top 2 layers)	14,178,820	12.95%	92.36	92.35	13.1
Full fine-tuning	109,485,316	100.0%	93.01	93.00	16.6

Single run per configuration with a fixed seed. Re-running with a different seed moves these numbers by roughly +/- 1-3 points, so gaps smaller than that are noise rather than findings.

Labels

• 0 World
• 1 Sports
• 2 Business
• 3 Sci/Tech

How to use

from transformers import pipeline

pipe = pipeline("text-classification", model="joseeangel/bert-base-uncased-agnews-topic")

Limitations and bias

Trained on 2004-era news wire, so topic drift is real: modern entities and technology vocabulary are under-represented, and the Business / Sci-Tech boundary is the model's weakest. Texts far longer than a headline and lead are truncated at 128 subwords. English only.

The model inherits whatever social bias is present in BERT's pretraining corpus (BooksCorpus + English Wikipedia) and in the task dataset above.

References

• Devlin, Chang, Lee & Toutanova (2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv:1810.04805 - section 5.3 is the feature-based vs fine-tuning comparison this work mirrors.
• Zhang, Zhao & LeCun (2015), Character-level Convolutional Networks for Text Classification
• HuggingFace Transformers, fine-tune a pretrained model.