---
task_categories:
- text-classification
- feature-extraction
language:
- en
---

# Text Quality Meta-Analysis Dataset

## Dataset Summary

The Text Quality Meta-Analysis Dataset is a comprehensive collection of sentences with associated quality metrics derived from multiple sources and methods. It combines text from various sources with quality scores from different models to create a thorough assessment of sentence quality. This dataset is an expanded and streamlined version of [agentlans/text-quality](https://huggingface.co/datasets/agentlans/text-quality).

In this context, "quality" refers to legible English sentences that are not spam and contain useful information. It does not necessarily indicate grammatical or factual correctness.

## Supported Tasks

This dataset is suitable for:

- Sentence quality prediction
- Text classification
- Natural language processing model evaluation
- Linguistic analysis of sentence structure and quality

## Dataset Structure

### Data Instances

Each instance contains:

- `text` (string): The sentence text
- `fineweb` (float): Quality score from [HuggingFaceFW/fineweb-edu-classifier](https://huggingface.co/HuggingFaceFW/fineweb-edu-classifier) after transformation
- `nvidia` (float): Quality score from [nvidia/quality-classifier-deberta](https://huggingface.co/nvidia/quality-classifier-deberta) after transformation
- `length` (float): Length of the sentence in characters after transformation
- `quality` (float): Composite quality score derived from other metrics

### Data Splits

The dataset is divided into two main files:

1. `all.csv.gz`: Contains all processed sentences
2. `train.csv.gz`: A balanced subset for training, with up to 10 000 sentences per quality group

## Dataset Creation

### Source Data

Sentences were collected from multiple sources:

- [allenai/c4](https://huggingface.co/datasets/allenai/c4)
- [HuggingFaceFW/fineweb-edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu)
- [monology/pile-uncopyrighted](https://huggingface.co/datasets/monology/pile-uncopyrighted)

### Annotations

Quality scores were generated using:

- HuggingFaceFW/fineweb-edu-classifier (logits)
- nvidia/quality-classifier-deberta (difference of logits for "High" and "Low" classes)

### Quality Score Derivation

A composite quality score was derived through the following steps:

1. Principal Component Analysis (PCA) was performed on the normalized "fineweb" and "nvidia" scores.
2. The first principal component was extracted as an initial measure of quality.
3. This quality measure was then adjusted for sentence length using robust linear regression (rlm function from the MASS package).
4. The adjusted quality scores were scaled to z-scores produce the final quality metric.

### Personal and Sensitive Information

No personal or sensitive information is included in this dataset.

## Considerations for Using the Data

### Social Impact of Dataset

This dataset can be used to improve natural language processing models and text quality assessment tools. However, users should be aware of potential biases in the source data and scoring models. These potential biases should be considered when interpreting results or developing applications based on this data.

### Known Limitations

- The composite quality score is derived from limited sources and may not capture all aspects of text quality.
- The dataset focuses only on English language sentences.
- It doesn't differentiate between true and false sentences, only their potential information value.
- Potential biases in the source data and scoring models may affect the overall quality assessment.