metadata
language:
- en
license: apache-2.0
task_categories:
- text-classification
tags:
- data-cleaning
- corpus-filtering
- text-quality
- ultrax
size_categories:
- 1M<n<10M
UltraX Line Classification Dataset
This dataset provides line-level annotations extracted from large English corpora to determine whether individual lines should be kept, deleted, or revised for data cleaning and curation.
This dataset is inspired by the UltraX project but is not officially affiliated with it. No proprietary artifacts from the original project are republished here.
Methodology
- Data Collection: 50,000 rows were sampled from each configuration of openbmb/UltraX-Preview.
- Line Comparison: Raw and cleaned text versions from each row were compared to classify individual lines into mutually exclusive categories:
- Keep: Lines retained in the final clean text.
- Delete: Lines removed during the cleaning process.
- Edit: Lines that underwent revision between the raw and clean versions.
- Splitting: The dataset uses a stratified split into train (80%), validation (10%), and test (10%).
Class Distribution
| Label | Train | Validation | Test | Total |
|---|---|---|---|---|
| Delete | 644,545 | 80,568 | 80,568 | 805,681 |
| Edit | 46,964 | 5,871 | 5,871 | 58,706 |
| Keep | 3,518,387 | 439,798 | 439,799 | 4,397,984 |
| Total | 4,209,896 | 526,237 | 526,238 | 5,262,371 |
Limitations
- Class Imbalance: The dataset is heavily imbalanced, reflecting real-world data cleaning distributions (the vast majority of lines are kept, fewer are deleted, and only a small fraction are edited).
- Context Dependency: Annotations are performed at the line level, which may occasionally lack broader document context (e.g., a specific line might serve as a valid title in one document but an irrelevant footnote in another).
- Formatting Artifacts: Potential structural quirks or formatting artifacts may persist from the raw text inputs.
License
Distributed under the Apache 2.0 License.