|
|
--- |
|
|
license: apache-2.0 |
|
|
task_categories: |
|
|
- text-classification |
|
|
- text-generation |
|
|
--- |
|
|
# MySuperDataset |
|
|
<!-- markdownlint-disable first-line-h1 --> |
|
|
<!-- markdownlint-disable html --> |
|
|
<!-- markdownlint-disable no-duplicate-header --> |
|
|
|
|
|
<div align="center"> |
|
|
<img src="figures/dataset_overview.png" width="60%" alt="MySuperDataset" /> |
|
|
</div> |
|
|
<hr> |
|
|
|
|
|
<div align="center" style="line-height: 1;"> |
|
|
<a href="LICENSE" style="margin: 2px;"> |
|
|
<img alt="License" src="figures/license_badge.png" style="display: inline-block; vertical-align: middle;"/> |
|
|
</a> |
|
|
</div> |
|
|
|
|
|
## 1. Introduction |
|
|
|
|
|
MySuperDataset is a comprehensive multi-domain text corpus designed for training and evaluating language models. This dataset has been carefully curated through multiple iterations to ensure high quality, diversity, and representativeness across various domains including science, technology, arts, and everyday conversations. |
|
|
|
|
|
<p align="center"> |
|
|
<img width="80%" src="figures/quality_chart.png"> |
|
|
</p> |
|
|
|
|
|
The dataset includes over 10 million samples spanning 15 different quality metrics. Through our iterative curation process, we have significantly improved data quality while maintaining diversity. The latest version shows a 25% improvement in overall quality scores compared to the initial release. |
|
|
|
|
|
Our curation process focuses on removing harmful content, deduplication, and ensuring balanced representation across topics and demographics. |
|
|
|
|
|
## 2. Quality Metrics |
|
|
|
|
|
### Comprehensive Quality Assessment |
|
|
|
|
|
<div align="center"> |
|
|
|
|
|
| | Metric | Baseline | Version1 | Version2 | MySuperDataset | |
|
|
|---|---|---|---|---|---| |
|
|
| **Content Quality** | Factual Accuracy | 0.720 | 0.745 | 0.768 | 0.829 | |
|
|
| | Grammar Score | 0.850 | 0.872 | 0.890 | 0.879 | |
|
|
| | Coherence | 0.780 | 0.795 | 0.812 | 0.873 | |
|
|
| **Diversity Metrics** | Topic Coverage | 0.650 | 0.680 | 0.705 | 0.773 | |
|
|
| | Vocabulary Richness | 0.720 | 0.738 | 0.755 | 0.806 | |
|
|
| | Style Variety | 0.680 | 0.702 | 0.720 | 0.768 | |
|
|
| | Source Diversity | 0.590 | 0.625 | 0.658 | 0.743 | |
|
|
| **Safety & Ethics** | Toxicity Filter | 0.920 | 0.935 | 0.948 | 0.963 | |
|
|
| | Bias Mitigation | 0.780 | 0.805 | 0.825 | 0.847 | |
|
|
| | Privacy Protection | 0.850 | 0.872 | 0.888 | 0.907 | |
|
|
| | Copyright Compliance | 0.910 | 0.925 | 0.938 | 0.960 | |
|
|
| **Technical Quality** | Deduplication Rate | 0.880 | 0.905 | 0.922 | 0.935 | |
|
|
| | Format Consistency | 0.820 | 0.845 | 0.865 | 0.880 | |
|
|
| | Encoding Quality | 0.950 | 0.962 | 0.970 | 0.970 | |
|
|
| | Metadata Completeness | 0.750 | 0.778 | 0.800 | 0.828 | |
|
|
|
|
|
</div> |
|
|
|
|
|
### Overall Quality Summary |
|
|
MySuperDataset demonstrates exceptional quality across all evaluated metrics, with particularly strong performance in safety filtering and content quality dimensions. |
|
|
|
|
|
## 3. Dataset Access & API |
|
|
We provide direct dataset access through our platform and API endpoints. Please visit our official website for documentation and access tokens. |
|
|
|
|
|
## 4. How to Use |
|
|
|
|
|
Please refer to our documentation for detailed usage instructions. |
|
|
|
|
|
### Loading the Dataset |
|
|
```python |
|
|
from datasets import load_dataset |
|
|
|
|
|
dataset = load_dataset("MySuperDataset") |
|
|
``` |
|
|
|
|
|
### Data Format |
|
|
Each sample contains: |
|
|
- `text`: The main text content |
|
|
- `domain`: The topic domain classification |
|
|
- `quality_score`: Pre-computed quality score |
|
|
- `source`: Original data source identifier |
|
|
|
|
|
### Recommended Usage |
|
|
We recommend filtering samples with quality_score >= 0.8 for training language models. |
|
|
|
|
|
## 5. License |
|
|
This dataset is released under the [Apache 2.0 License](LICENSE). Commercial use is permitted with proper attribution. |
|
|
|
|
|
## 6. Contact |
|
|
For questions or issues, please open a GitHub issue or contact us at data@mysuperdataset.ai. |
|
|
|
