inokusan/human_ai_text_classification

---
pretty_name: Human AI Text Classification
language:
- en
license: other
task_categories:
- text-classification
task_ids:
- binary-classification
size_categories:
- 100K<n<1M
tags:
- ai-generated-text
- human-vs-ai
- text-classification
- llm-detection
- english
dataset_info:
features:
- name: text
dtype: string
- name: label
dtype: int64
splits:
- name: train
num_bytes: 97511886
num_examples: 90648
- name: test
num_bytes: 24329778
num_examples: 22662
download_size: 68869429
dataset_size: 121841664
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
---

# Human AI Text Classification

## Dataset Summary

`human_ai_text_classification` is a binary text classification dataset for distinguishing human-written text from AI-generated text.

It was created by combining three public datasets, standardizing them into a common schema, balancing the class labels, removing duplicate texts, and performing a stratified 80/20 train-test split.

Labels:
- `0` = human-written text
- `1` = AI-generated text

## Dataset Structure

### Data Fields

- `text`: the input text
- `label`: binary class label
- `0` for human-written text
- `1` for AI-generated text

### Splits

- `train`: 90,648 rows
- `test`: 22,662 rows

### Label Distribution

This final dataset is globally balanced:

- Total rows: `113,310`
- Human (`0`): `56,655`
- AI (`1`): `56,655`

Split-level balance:

- Train: `45,324` human, `45,324` AI
- Test: `11,331` human, `11,331` AI

## Dataset Creation

### Source Datasets

This dataset was built from the following original sources:

1. `NicolaiSivesind/human-vs-machine`
https://huggingface.co/datasets/NicolaiSivesind/human-vs-machine

2. `thedrcat/daigt-v2-train-dataset`
https://www.kaggle.com/datasets/thedrcat/daigt-v2-train-dataset

3. `shahxeebhassan/human_vs_ai_sentences`
https://huggingface.co/datasets/shahxeebhassan/human_vs_ai_sentences

### Processing Steps

The dataset was created with the following pipeline:

1. Load the three source datasets.
2. Standardize columns to `text` and `label`.
3. Standardize labels to:
- `0` for human
- `1` for AI
4. Sample each source to keep it internally balanced:
- `NicolaiSivesind/human-vs-machine`: `20,000` human + `20,000` AI
- `thedrcat/daigt-v2-train-dataset`: `17,497` human + `17,497` AI
- `shahxeebhassan/human_vs_ai_sentences`: `20,000` human + `20,000` AI
5. Merge all sampled subsets.
6. Remove duplicate texts using `drop_duplicates(subset=["text"])`.
7. Rebalance globally after deduplication to preserve exact class balance.
8. Shuffle the full dataset.
9. Perform a stratified 80/20 train-test split.

### Resulting Dataset Size

- Before deduplication: `114,994`
- After deduplication: `113,886`
- Final balanced size: `113,310`

## Intended Use

This dataset is intended for:
- training baseline AI-text detectors
- benchmarking binary human-vs-AI text classification
- experiments on generalization across mixed-source human and machine-generated writing

## Limitations

- The dataset combines multiple source datasets with different collection methods and writing styles.
- It should not be treated as a universal detector for all LLM-generated text.
- Label quality depends on the correctness of the original source datasets.
- Some source datasets are themselves aggregated from earlier datasets or competition resources.

## License and Attribution

This dataset is a derived compilation of multiple public datasets.
Please review the original dataset pages for the applicable licenses, usage terms, and attribution requirements before reuse or redistribution.

Original dataset references:

- `NicolaiSivesind/human-vs-machine`
https://huggingface.co/datasets/NicolaiSivesind/human-vs-machine

- `thedrcat/daigt-v2-train-dataset`
https://www.kaggle.com/datasets/thedrcat/daigt-v2-train-dataset

- `shahxeebhassan/human_vs_ai_sentences`
https://huggingface.co/datasets/shahxeebhassan/human_vs_ai_sentences

## Loading the Dataset

```python
from datasets import load_dataset

dataset = load_dataset("inokusan/human_ai_text_classification")
print(dataset)
```