MultiSocial

MultiSocial is a dataset (described in a paper) for multilingual (22 languages) machine-generated text detection benchmark in social-media domain (5 platforms). It contains 472,097 texts, of which about 58k are human-written and approximately the same amount is generated by each of 7 multilingual large language models by using 3 iterations of paraphrasing. The dataset has been anonymized to minimize amount of sensitive data by hiding email addresses, usernames, and phone numbers.

If you use this dataset in any publication, project, tool or in any other form, please, cite the paper.

Disclaimer

Due to data source (described below), the dataset may contain harmful, disinformation, or offensive content. Based on a multilingual toxicity detector, about 8% of the text samples are probably toxic (from 5% in WhatsApp to 10% in Twitter). Although we have used data sources of older date (lower probability to include machine-generated texts), the labeling (of human-written text) might not be 100% accurate. The anonymization procedure might not successfully hiden all the sensitive/personal content; thus, use the data cautiously (if feeling affected by such content, report the found issues in this regard to dpo[at]kinit.sk). The intended use if for non-commercial research purpose only.

Data Source

The human-written part consists of a pseudo-randomly selected subset of social media posts from 6 publicly available datasets:

1. Telegram data originated in Pushshift Telegram, containing 317M messages (Baumgartner et al., 2020). It contains messages from 27k+ channels. The collection started with a set of right-wing extremist and cryptocurrency channels (about 300 in total) and was expanded based on occurrence of forwarded messages from other channels. In the end, it thus contains a wide variety of topics and societal movements reflecting the data collection time.

2. Twitter data originated in CLEF2022-CheckThat! Task 1, containing 34k tweets on COVID-19 and politics (Nakov et al., 2022, combined with Sentiment140, containing 1.6M tweets on various topics (Go et al., 2009).

3. Gab data originated in the dataset containing 22M posts from Gab social network. The authors of the dataset (Zannettou et al., 2018) found out that “Gab is predominantly used for the dissemination and discussion of news and world events, and that it attracts alt-right users, conspiracy theorists, and other trolls.” They also found out that hate speech is much more prevalent there compared to Twitter, but lower than 4chan's Politically Incorrect board.

4. Discord data originated in Discord-Data, containing 51M messages. This is a long-context, anonymized, clean, multi-turn and single-turn conversational dataset based on Discord data scraped from a large variety of servers, big and small. According to the dataset authors, it contains around 0.1% of potentially toxic comments (based on the applied heuristic/classifier). 

5. WhatsApp data originated in whatsapp-public-groups, containing 300k messages (Garimella & Tyson, 2018). The public dataset contains the anonymised data, collected for around 5 months from around 178 groups. Original messages were made available to us on request to dataset authors for research purposes.

From these datasets, we have pseudo-randomly sampled up to 1300 texts (up to 300 for test split and the remaining up to 1000 for train split if available) for each of the selected 22 languages (using a combination of automated approaches to detect the language) and platform. This process resulted in 61,592 human-written texts, which were further filtered out based on occurrence of some characters or their length, resulting in about 58k human-written texts.

The machine-generated part contains texts generated by 7 LLMs (Aya-101, Gemini-1.0-pro, GPT-3.5-Turbo-0125, Mistral-7B-Instruct-v0.2, opt-iml-max-30b, v5-Eagle-7B-HF, vicuna-13b). All these models were self-hosted except for GPT and Gemini, where we used the publicly available APIs. We generated the texts using 3 paraphrases of the original human-written data and then preprocessed the generated texts (filtered out cases when the generation obviously failed).

The dataset has the following fields:

• 'text' - a text sample,

• 'label' - 0 for human-written text, 1 for machine-generated text,

• 'multi_label' - a string representing a large language model that generated the text or the string "human" representing a human-written text,

• 'split' - a string identifying train or test split of the dataset for the purpose of training and evaluation respectively,

• 'language' - the ISO 639-1 language code identifying the detected language of the given text,

• 'length' - word count of the given text,

• 'source' - a string identifying the source dataset / platform of the given text,

• 'potential_noise' - 0 for text without identified noise, 1 for text with potential noise.

ToDo Statistics (under construction)