Assignment 1 / Text Track

Text Classification

20 Newsgroups — fine-tuned BERT vs. BiLSTM with frozen BERT embeddings, topped with a learned ensemble meta-learner.

Track Setup

Component	Selection
Dataset	20 Newsgroups — 11,314 train / 7,532 test / 20 classes
Tokenizer	bert-base-uncased — WordPiece, vocab 30,522, max_length=256
Baseline	BiLSTM with frozen BERT embeddings + attention pooling — 7.4 M trainable params
Advanced	Fine-tuned bert-base-uncased (AutoModelForSequenceClassification) — 109.5 M params, AMP FP16
Ensemble	Learned meta-MLP over pre-computed logits from both models — 7,340 trainable params
Best accuracy	70.23% (BERT) on 7,532 test samples — BiLSTM 68.00%, Ensemble 69.48%
Experiment tracking	Weights & Biases — 3 runs (BERT, BiLSTM, Ensemble)
GPU	NVIDIA RTX 3060 Laptop (6 GB VRAM) — CUDA 12.1, PyTorch 2.5.1

Track Pages

Open each section to see the full report for the Text track.

Section

Dataset EDA

20-class distribution, word-count profiles, BERT tokenization decisions, and empty-text handling for 20 Newsgroups.

Open Dataset EDA

Section

Model Backbone

Full specs for the fine-tuned BERT classifier, BiLSTM with frozen BERT embeddings, and the ensemble meta-MLP — with parameter counts and design rationale.

Open Model Backbone

Section

Methodology

End-to-end pipeline, hyperparameter tables, AMP + gradient-accumulation strategy, regularisation, and learning curves for all three models.

Open Methodology

Section

Evaluation Results

Test accuracy (BERT 70.23%), macro F1, per-class F1 bars, confusion matrices, and ROC AUC — with error-analysis and next-step recommendations.

Open Evaluation Results