BART-MNLI CNN News Zero-Shot Classifier
A BART-MNLI model adapted to CNN news for topic classification through the Hugging Face zero-shot pipeline.
Timeline
Published to Hugging Face on July 12, 2023, after data prep, fine-tuning, and model-card documentation.
Role & Scope
Prepared CNN news data, fine-tuned the model, and packaged the weights, tokenizer, and model card for reuse.
Team Setup
Co-developed with Faycal Boubekri (as listed in the model card), with shared validation on category behavior and outputs.
Context
Problem: News categorization needed a reusable model that could score candidate topic labels in a standard inference workflow.
Solution: Fine-tuned facebook/bart-large-mnli on CNN articles and published the model, tokenizer, and a Transformers inference example.
Impact: The model card reports 94% F1 and accuracy on the CNN test split; weights are available for local inference.
Implementation Details
Architecture
BART-Large-MNLI retains contradiction, neutral, and entailment outputs. The zero-shot pipeline scores article and candidate-label pairs using the entailment prediction; published weights use safetensors.
Data & Evaluation
The card documents one training epoch with 256-token inputs and CNN test evaluation at 128 tokens. Reported scores describe that test setup; they do not establish performance on unseen domains or arbitrary labels.
Production & Ops
Published under Apache-2.0 with a Transformers loading example. The Hub provides downloadable artifacts; a hosted inference service is not bundled with the project.
Lessons Learned
Candidate-label wording and input truncation deserve explicit evaluation when adapting an NLI model to a specific news domain.
Links
Last verified: September 1, 2026
Public Evidence
Last verified: September 1, 2026