Core Concepts
Natural Language Inference enables efficient universal classification tasks without fine-tuning.
Abstract
The content discusses the efficiency of using Natural Language Inference (NLI) for building universal classifiers without the need for fine-tuning. It explains the concept, provides a step-by-step guide with reusable Jupyter notebooks, and shares insights from training a universal classifier on various datasets. The paper highlights the advantages of NLI-based classifiers over generative Large Language Models (LLMs) and emphasizes the importance of data preprocessing, cleaning, hypothesis formulation, training, evaluation, and visualization of results.
Introduction
Generative models' rise in academia.
Importance of resource-efficient universal models.
NLI as a universal task
Definition and examples of Natural Language Inference.
Transformation into binary entailment vs. not-entailment task.
Building a universal classifier
Dataset selection and harmonization.
Automatic data cleaning using CleanLab library.
Hypothesis formulation and NLI formatting.
Training and evaluation
Use of pre-trained transformer models like DeBERTaV3.
Evaluation metrics like balanced accuracy.
Visualisation and interpretation of results
Performance comparison between NLI-only model and mixed-data model.
Reusing models and code
Recommendations for downstream use or fine-tuning.
Limitations
Data diversity limitations, noise in data, computational overhead for high-class tasks.
Stats
"Our new classifier improves zero-shot performance by 9.4%."
"Parts of the code we share has been used to train our older zero-shot classifiers that have been downloaded more than 55 million times via the Hugging Face Hub as of December 2023."
Quotes
"We need to raise tariffs"
"It is about economy"
"Our armed forces keep us safe"