AUGFOX: An augmented dataset of argumentative texts written by humans and machines

The AUGFOX dataset consists of 215.600 argumentative texts written by humans and various open and closed-source language models. It is an augmented version of OUTFOX (Koike et al. 2024) using recent open-weight models and ChatGPT versions. `augfox.csv` contains the augmented dataset with genericity scores, and `rst.zip` contains the RST parses obtained using the DMRST parser (Liu et al. 2021). Fields unique_id: id identifying each unique text. model: model name.text: each generated (or human-written) text. clause2labels: clause labels.genericity_score: the genericity score.split: data split.*local_idx: id of each text inside a combination of model and split.global_idx: id of text across models and splits.attacks: attacked texts.*problem_statement: problem statement.* prompt: the prompt used to generated each text.* *: propagated from Outfox. Model distribution model train valid test total apertus-70b-instruct 14400 500 500 15400 apertus-8b-instruct 14400 500 500 15400 chatgpt-3-5 14400 500 500 15400 chatgpt-4o 14400 500 500 15400 chatgpt-5-2 14400 500 500 15400 flan-t5-xxl 14400 500 500 15400 gemma-3-27b-it 14400 500 500 15400 gemma-3-12b-it 14400 500 500 15400 gemma-3-4b-it 14400 500 500 15400 ministral-3-14b-instruct 14400 500 500 15400 ministral-3-8b-instruct 14400 500 500 15400 ministral-3-3b-instruct 14400 500 500 15400 text_davinci_003 14400 500 500 15400 human 14400 500 500 15400 215600