File size: 2,906 Bytes
4a70d98
 
 
 
 
 
 
 
37229f2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4a70d98
37229f2
4a70d98
37229f2
4a70d98
37229f2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
---
license: cc
task_categories:
- text-classification
- feature-extraction
language:
- en
---
# Text Quality Assessment Dataset

## Overview

This dataset is designed to assess text quality robustly across various domains for NLP and AI applications. It provides a composite quality score based on multiple classifiers, offering a more comprehensive evaluation of text quality beyond educational domains.

## Dataset Details

- **Size**: 100,000 sentences
- **Source**: 20,000 sentences from each of 5 different datasets
  - [allenai/c4](https://huggingface.co/datasets/)
  - [HuggingFaceFW/fineweb-edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu)
  - [monology/pile-uncopyrighted](https://huggingface.co/datasets/monology/pile-uncopyrighted)
  - [agentlans/common-crawl-sample](https://huggingface.co/datasets/agentlans/common-crawl-sample)
  - [agentlans/wikipedia-paragraphs](https://huggingface.co/datasets/agentlans/wikipedia-paragraphs)

## Features

The quality scores of each text were assessed using
  - [HuggingFaceFW/fineweb-edu-classifier](https://huggingface.co/HuggingFaceFW/fineweb-edu-classifier)
  - [nvidia/quality-classifier-deberta](https://huggingface.co/nvidia/quality-classifier-deberta)

1. **Text Length**: 
   - Measured in characters
   - Box-Cox transformed

2. **Fineweb-edu Classifier Score**:
   - Raw logits
   - Yeo-Johnson transformed

3. **NVIDIA Quality Score**:
   - Logits of "High" quality level - logits of "Low" quality level

5. **Composite Quality Score**:
   - First principal component of fineweb-edu and NVIDIA scores
   - Adjusted for length using linear regression with the transformed text length

## Key Insights

- Fineweb-edu and NVIDIA scores show weak correlation
- Composite quality score correlates with both individual scores
- Clear quality differences observed across the 5 source datasets

**Figure 1**: Correlation between individual scores (fineweb-edu and NVIDIA) and the composite quality score. Each point represents a single row of text.
<img src="https://huggingface.co/datasets/agentlans/text-quality/resolve/main/CorrelationPlot.png" alt="Quality score scatterplot" width="50%"/>

**Figure 2**: Distribution of quality scores across the five source datasets, highlighting quality differences
<img src="https://huggingface.co/datasets/agentlans/text-quality/resolve/main/QualityDistribution.png" alt="Quality score scatterplot" width="75%"/>

## Applications

- Benchmarking text quality across various domains
- Training robust text quality assessment models
- Analyzing dataset quality for diverse NLP tasks

## Limitations

- Based on existing classifiers, may inherit their biases
- The current quality definition may not capture all aspects of text quality

## Ethics and Privacy

- No personal information is included in the dataset
- Users should appropriately credit the source datasets when using this compilation