CHALLENGE-IT

Twelfth Italian Conference on Computational Linguistics, September 14–16, 2026, Palermo, Italy. Last updated: 2026-09-11.

Italian benchmark for LLM-generated text detection with 360,438 samples from four Italian corpora (CHANGE-IT, C4-IT, PaISà), generated by 14 LLMs. The test split is a hard negative subset of 9,294 examples simultaneously difficult for both a RoBERTa-based and an EuroBERT-based classifier — all evaluated systems score AUC 0.39–0.56 on this subset (near random chance).

Splits

Split	Rows	Notes
train	298,472	Training set
validation	52,672	Validation set
test	9,294	Hard negative subset

Usage

from datasets import load_dataset

ds = load_dataset("alessandropisent/CHALLENGE-IT")
hard_test = ds["test"]

Citation

@inproceedings{pisent2026challengeit,
  title     = {CHALLENGE-IT: A Multi-Source Italian Benchmark with Hard Negatives
               for LLM-Generated Text Detection},
  author    = {Pisent, Alessandro and Silvestri, Fabrizio},
  booktitle = {Proceedings of the Twelfth Italian Conference on
               Computational Linguistics (CLiC-it 2026)},
  year      = {2026},
  address   = {Palermo, Italy}
}