HCVAR: PARAPHRASED, ORIGINAL (GPT,LLAMA,MISTRAL)

Dataset Overview:------------------

We have randomly sampled 2000 samples for each of the following domains from the HCVAR data.

Essay

News

Q&A

Review

It originally contained machine-generated text by GPT. We additionally generated using Llama and Mistral as well.

We paraphrased a portion of the AI-generated text using a Pegasus-based paraphraser. (details are given below)

Paraphrasing Information:--------------------------

50% of the AI-generated text in the dataset is paraphrased, while the other half remains unchanged as originally generated.

The selection of AI text for paraphrasing is random.

Dataset Structure:------------------Each file in the dataset contains the following required columns:

original_text: The non-paraphrased text.

label:   - 1: Human-generated text   - 0: AI-generated text

is_selected:   - 1: The AI text was selected for paraphrasing   - 0: The text was not selected for paraphrasing

text: The paraphrased version of the AI-generated text (if applicable), non-paraphrased otherwise.