README.md

---
license: odc-by
task_categories:
- text-classification
- text-generation
- feature-extraction
- sentence-similarity
language:
- en
---
# High-Quality English Sentences

## Dataset Description

This dataset contains a collection of high-quality English sentences sourced from [C4](https://huggingface.co/datasets/allenai/c4) and [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) (*not* [FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu)). The sentences have been carefully filtered and processed to ensure quality and uniqueness.

"High-quality" means they're legible English and not spam, although they may still have spelling and grammar errors.

## Source Data

Before filtering:
- **C4:** 1 million sentences
- **FineWeb:** 1 million sentences

## Data Processing

1. **Initial Sentence Filtering:**
   - Quality score > 0.5 with [agentlans/deberta-v3-xsmall-quality](https://huggingface.co/agentlans/deberta-v3-xsmall-quality) classifier
   - Length >= 20 characters

2. **Additional Filtering:**
   - Removed sentences not starting with a capital letter
   - Removed sentences with unmatched parentheses or brackets

3. **Deduplication:**
   - Combined sentences from both sources
   - Filtered for unique sentences (exact match)

4. **Train-Test Split:**
   - Train set: 90% of data
   - Test set: 10% of data

## Data Fields

No data field. Just one sentence per line as a string.

## Data Splits

| Split   | Number of Sentences |
|---------|---------------------|
| Train   | 1 534 699           |
| Test    | 170 522             |
| Total   | 1 705 221           |

## Dataset Creation

- **Curation Rationale:** To create a high-quality dataset of diverse English sentences for various NLP tasks.
- **Source Contributions:** Equal initial contributions from C4 and FineWeb, with subsequent filtering and processing.

## Considerations for Using the Data

- The dataset deliberately uses FineWeb and C4 instead of FineWeb-Edu in order to broaden beyond educational and academic content. However, it's still biased towards those genres.
- Titles and citations are considered "sentences".
- The filtering process may have introduced some biases, such as favouring more formal writing styles and omitting lists due to the capitalization requirement.