---
language:
- en
pretty_name: "Text Quality Dataset"
tags:
- text-quality
- classification
license: odc-by
task_categories:
- text-classification
---

# Text Quality Dataset

## Overview
This dataset contains **100,000 rows** sampled from the `allenai/c4` English split, annotated with various text quality classifiers.

## Methods

Classifiers used:

| Label         | Model                                                                 | Method         |
|---------------|-----------------------------------------------------------------------|----------------|
| fineweb2hq    | [agentlans/multilingual-e5-small-fineweb2hq-vs-c4-classifier](https://huggingface.co/agentlans/multilingual-e5-small-fineweb2hq-vs-c4-classifier) | Logit difference |
| finewebedu    | [HuggingFaceFW/fineweb-edu-classifier](https://huggingface.co/HuggingFaceFW/fineweb-edu-classifier)           | Raw logits     |
| gneiss       | [ibm-granite/GneissWeb.Quality_annotator](https://huggingface.co/ibm-granite/GneissWeb.Quality_annotator)     | FastText       |
| nemo         | [nvidia/nemocurator-fineweb-nemotron-4-edu-classifier](https://huggingface.co/nvidia/nemocurator-fineweb-nemotron-4-edu-classifier) | Raw logits     |
| nvidia       | [nvidia/quality-classifier-deberta](https://huggingface.co/nvidia/quality-classifier-deberta)                   | Logit difference |
| ultrafineweb | [openbmb/Ultra-FineWeb-classifier](https://huggingface.co/openbmb/Ultra-FineWeb-classifier)                     | FastText       |
| uvp          | [agentlans/bge-small-en-v1.5-ultrafineweb-vs-pile-classifier](https://huggingface.co/agentlans/bge-small-en-v1.5-ultrafineweb-vs-pile-classifier) | Logit difference |

For FastText-based classifiers, the classifiers' output probabilities were converted to logits log(p).
The **quality score** for those classifiers were computed as the difference of logits between the high-quality and low-quality classes.

### Overall score computation
- Scores from all classifiers were centreed and scaled.
- Principal Components Analysis (PCA) was applied.
- The first principal component (PC1) was normalized to z-scores (mean 0, standard deviation 1)
- The z-score is taken as the overall quality score.

For convenience, the dataset is split into an **80% training set** and a **20% testing set**.

## Results

![Scatterplot of quality scores from different classifiers](Scatterplot.svg)

- The **lower triangle** shows pairwise density plots of classifier scores.
- The **diagonal** presents the distribution of each classifier's scores.
- The **upper triangle** displays correlations between pairs of classifiers.

- The classifiers' scores show moderate to strong correlations, except for the Nvidia classifier, which is less correlated.
- The custom-trained classifiers often give bimodal distributions instead of smoothly varying values.
- Despite the above, the overall score correlates well with each individual classifier's quality score.

## License
This dataset is licensed under the **[Open Data Commons Attribution License (ODC-BY)](https://opendatacommons.org/licenses/by/)**.