Dataset · ming9999
LLM text authorship attribution benchmark
ming9999/maabench
English dataset for classifying generated texts across five LLM families with domain and version shift testing.
What's inside
A set of English text documents (between 100 thousand and 1 million examples in jsonl format) generated by five model families: gpt, claude, gemini, mistral, and llama. The data covers academic texts, comments, news, and wiki articles, split into a baseline sample and distribution-shift tests (new domains and fresh model versions).
What you can do with it
- Training classifiers for attributing AI texts to specific LLM families
- Evaluating the robustness of synthetic text detectors to new model versions and unfamiliar topics
- Tasks
- Text classification
- Languages
- English
- Size
- 100K–1M rows
- Downloads this month
- 294
- Likes
- 7
- Access
- on request on Hugging Face
How to load it
from datasets import load_dataset
ds = load_dataset("ming9999/maabench")