Datasets:

mohamedmady
/

HC3-Gemini-Flash-Responses

license: mit
task_categories:
  - text-classification
language:
  - en
tags:
  - ai-generated-text-detection
  - generator-shift
size_categories:
  - 10K<n<100K

Single-class dataset. Every row is AI-generated (Is_AI = 1). It is designed to measure generator shift and must be paired with human text, e.g. the human side of HC3 PLUS, to form a balanced detection benchmark.

HC3 Gemini 2.0 Flash Responses

Dataset Description

This dataset contains {len(df_upload):,} AI-generated text samples produced by Google Gemini 2.0 Flash in response to questions from the HC3 (Human ChatGPT Comparison Corpus) benchmark.

It was constructed to evaluate AI-generated text detectors on out-of-distribution generator transfer — specifically, testing whether detectors trained on other generators (ChatGPT, LLaMA, etc.) can detect Gemini-generated text without retraining.

Usage

from datasets import load_dataset
ds = load_dataset("mohamedmady/HC3-Gemini-Flash-Responses")

Dataset Details

Property	Value
Samples	{len(df_upload):,}
Generator	Gemini 2.0 Flash
Source questions	HC3 benchmark
Domain	Question Answering (QA)
Language	English
Label	All AI-generated (label=1)

Column Schema

Column	Type	Description
`text`	string	AI-generated response text
`label`	ClassLabel	1 = AI-generated, 0 = Human
`source`	string	Generator model (`gemini-2.0-flash`)
`domain`	string	Text domain (`qa`)
`split`	string	Dataset split (`test`)

Context and Citation

This dataset was created for the paper:

DeBERTa-ConPara: Robust AI-Generated Text Detection under Adversarial and Cross-Domain Conditions (EMNLP 2026 submission)

DeBERTa-ConPara v2.4-3k achieves 98.64% AI recall on this dataset under fixed-threshold evaluation (τ*=0.920), demonstrating strong cross-generator generalization to Gemini outputs without any Gemini training data.

@misc{{Mady25-GeminiHC3,
  author = {{Mady, Mohamed}},
  title  = {{HC3 Gemini 2.0 Flash Responses}},
  year   = {{2025}},
  url    = {{https://huggingface.co/datasets/mohamedmady/HC3-Gemini-Flash-Responses}},
  note   = {{Hugging Face dataset}}
}}

License

CC-BY-4.0. The source HC3 questions are from Hello-SimpleAI/HC3. Responses were generated by the authors using the Gemini 2.0 Flash API. """