NemoCurator Quality Classifier DeBERTa

Model Overview

This is a text classification model that can enable qualitative data annotation, creation of quality-specific blends and addition of metadata tags. The model classifies documents into one of three classes based on the quality of the document: "High", "Medium", "Low"

The model was trained using data annotated by human annotators, who considered quality factors such as content accuracy, clarity, coherence, grammar, depth of information, and overall usefulness of the document.

This model is used in the NVIDIA NeMo Curator as part of the qualitative filtering module.

Model Architecture

The model architecture is Deberta V3 Base Context length is 1024 tokens

Training (details)

Training data:

The training set is 22828 Common Crawl text samples, labeled as "High", "Medium", "Low". Here are some examples:

1. Input:

   Volunteering
   
   It's all about the warm, fuzzy feeling when you serve the community, without expectation of gain. Volunteering offers you the necessary experience and development skills to take forward with you, as you venture out to work with other people and apply what you learn, to achieve your career goals.
   
   HOW IT WORKS
   
   SEARCH
   
   BOOK NOW
   
   ENJOY THE SHOW
   
   GET A FREE QUOTE
   
   Planning your event ahead of time is the right move. Contact our experts and let us surprise you.
   

   Output: Low

2. Input:

   Sharapova has been in New Zealand since well before the New Year, preparing for her 2011 start and requested the opening day match to test her form. "My last tournament was over two months ago and it will be really good to get back playing again."
   
   "My priority since I have been here has been to adjust to time and conditions. I have had a couple of practices a day and think that has been really important."
   
   The three-time Grand Slam champion who once stood number one next plays Voracova after winning their only previous match in 2003.
   

   Output: High

How To Use This Model

Input

The model takes one or several paragraphs of text as input.

Example input:

Reasons to visit Thatta

Thatta is one of the most important cities of the province of Sindh, Pakistan. Historically it is the richest city. The sands of Thatta have seen many great men. It provided Alexander the Great and his troops a comfortable resting place before they moved further. It welcomed the Mughal Emperor Shah Jehan.

Output

The model outputs one of the 3 classes as the predicted quality for each input sample.

Example output:

Medium

How to use in NeMo Curator

The inference code is available on NeMo Curator's GitHub repository. Download the model.pth and check out this example notebook to get started.

Evaluation Benchmarks

Evaluation data

The evaluation data is a subset of training data where all three annotators agree on the label. It has 7128 samples.

Results

Accuracy score on evaluation set with 7128 samples - 0.8252

Confusion Matrix:

We verify that the predicted scores are indeed close to their ground truth, and are due to the noisy nature of the annotation.

Limitations

• Subjectivity in Quality: Quality assessment is inherently subjective and may vary among different annotators.

References

• https://arxiv.org/abs/2111.09543
• https://github.com/microsoft/DeBERTa

License

License to use this model is covered by the Apache 2.0. By downloading the public and release version of the model, you accept the terms and conditions of the Apache License 2.0. This repository contains the code for the domain classifier model.