| --- |
| language: |
| - en |
| tags: |
| - text-classification |
| - human-vs-machine |
| - synthetic-data |
| - data-quality |
| task_categories: |
| - text-classification |
| pretty_name: QualiText |
| size_categories: |
| - 100K<n<1M |
| license: cc0-1.0 |
| --- |
| |
| # QualiText |
|
|
| QualiText is a balanced English text-classification dataset for studying text |
| origin and text quality. Each example contains a `text` field and a `label` |
| field. The dataset has five labels with the same number of examples in each |
| class. |
|
|
| ## Labels |
|
|
| | Label | Description | |
| |---|---| |
| | `human` | Human-authored text from Wikipedia and 4chan. | |
| | `machine_generated` | Machine-generated text from the Qwen3.8-Max, GLM-5.2, and Kimi-K3 distillation corpus. | |
| | `corrupted` | Wikipedia text modified with deterministic word deletions, swaps, duplications, or typo-like edits. | |
| | `marketing` | Marketing-oriented text from FineWeb-Marketing. | |
| | `low_quality` | Spam, ham/phishing-email corpus examples, and synthetic text created by removing punctuation, lowercasing, and shuffling words. | |
|
|
| ## Sources |
|
|
| The source datasets are: |
|
|
| - [`wikimedia/wikipedia`](https://huggingface.co/datasets/wikimedia/wikipedia) |
| - [`fuzzy-g/4chan_pol_whole_ds`](https://huggingface.co/datasets/fuzzy-g/4chan_pol_whole_ds) |
| - [`r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation`](https://huggingface.co/datasets/r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation) |
| - [`marketeam/FineWeb-Marketing`](https://huggingface.co/datasets/marketeam/FineWeb-Marketing) |
| - [`seuun/spam-ham-phish-emails-latest`](https://huggingface.co/datasets/seuun/spam-ham-phish-emails-latest) |
|
|
| ## Data Format |
|
|
| ```text |
| text: string |
| label: ClassLabel or string |
| ``` |
|
|
| Example: |
|
|
| ```json |
| { |
| "text": "Example document text.", |
| "label": "human" |
| } |
| ``` |
|
|
| ## Intended Uses |
|
|
| QualiText can be used for: |
|
|
| - Training baseline text-origin classifiers |
| - Evaluating robustness to common text corruption |
| - Comparing text-quality classification strategies |
| - Prototyping data filtering and moderation models |
|
|
| It should not be used as the sole basis for deciding whether a person used an |
| AI system or whether content is trustworthy. |
|
|
| ## Limitations and Biases |
|
|
| - The labels represent dataset provenance and synthetic transformations, not |
| definitive proof of authorship. |
| - The human class contains platform-specific and encyclopedic writing styles. |
| - The marketing and low-quality classes may contain strong lexical shortcuts. |
| - Synthetic corruption does not represent every real-world form of corruption. |
| - The machine-generated class may reflect the style and artifacts of its |
| teacher models and prompts. |
| - Source datasets may contain offensive, private, copyrighted, or otherwise |
| sensitive material. Review examples before deployment. |
| - Performance may not generalize to languages, domains, or writing styles not |
| represented in the sources. |
|
|
| ## Licensing |
|
|
| QualiText is an aggregated dataset. The applicable license and usage terms of |
| each source dataset may differ and continue to apply to the corresponding |
| examples. Users are responsible for reviewing the source licenses and meeting |
| their attribution, privacy, copyright, and acceptable-use obligations. |
|
|
| ## Citation |
|
|
| If you use QualiText, cite the dataset repository where it is published and |
| also acknowledge the source datasets listed above. |
