petrematei/ro-political-ai
Dataset Card for "RO-Political-Texts" Dataset Description Dataset Summary RO-Political-AI is a collection of datasets for RoNLP, focused on detecting the difference between political texts written by humans, during the presidential elections from Romania (2024-2025), and synthetic texts generated with LLM models, as well as on studying linguistic characteristics such as slang, idioms, figurative expressions and "monkey business" behaviors in… See the full description on the dataset page: https://huggingface.co/datasets/petrematei/ro-political-ai.
Dataset Card for "RO-Political-Texts"
Table of Contents
- Considerations for Using the Data
- Social Impact of Dataset
- Discussion of Biases
- Other Known Limitations
- Ethical Considerations
Dataset Description
Dataset Summary
RO-Political-AI is a collection of datasets for RoNLP, focused on detecting the difference between political texts written by humans, during the presidential elections from Romania (2024-2025), and synthetic texts generated with LLM models, as well as on studying linguistic characteristics such as slang, idioms, figurative expressions and "monkey business" behaviors in political discourse on social networks.
The dataset includes:
- political posts, comments and replies (tweets) extracted from social media
- automatically generated completions, paraphrases, restylings (formal and informal) with OpenLLM-Ro instruct generative models (RoLlama2-7b-Instruct, RoLlama3-8b-Instruct, RoLlama3.1-8b-Instruct, RoMis-tral-7b-Instruct, RoGemma-7b-Instruct and RoGemma2-9b-Instruct)
Objectives:
- Detection of synthetic text
- Analysis of political polarization
- Robust evaluation of LLMs regarding underrepresented languages such as Romanian
Supported Tasks
- Synthetic text detection (Human vs AI)
- Toxicity classification
- Political discourse style analysis (slang, tone, bias)
- Study of linguistic differences between full human-written texts and texts containing AI-generated text fragments (mix of human writing + AI) -> AI Authorship
Languages
Romanian
Dataset Structure
Data Instances
An example of 'train' looks as follows.
{
"id": 1,
"human": "Dacă Nicușor ajunge președinte, o să fie vai și amar de noi. Promisiuni multe, dar fapte zero. Ne-a mai păcălit o dată la primărie, acum vine iar cu gogoși electorale.",
"RoLlama2-7b-Instruct-Completion": "Dacă Nicușor ajunge președinte, o să fie vai și amar de noi. Promisiuni multe, dar fapte zero. Ne-a mai păcălit o dată la primărie, acum vine iar cu gogoși electorale. Oamenii s-au săturat de vorbe frumoase fără rezultate. De data asta sper să nu mai punem botul la aceleași minciuni.",
"RoLlama2-7b-Instruct-Paraphrasing": "Dacă Nicușor ajunge președinte, ne așteaptă vremuri grele. Promite multe, dar nu face nimic. Ne-a mai înșelat când era la primărie, iar acum vine din nou cu aceleași povești electorale. Lumea s-a săturat de promisiuni goale. Sper ca de data asta să nu mai fim atât de naivi.",
"RoLlama2-7b-Instruct-Restyling-informal": "Dacă ajunge Nicușor președinte, suntem terminați. Numai vorbe și promisiuni degeaba. Ne-a mai păcălit o dată la primărie și acum vine iar cu aceleași vrăjeli. Oamenii s-au prins de filmul ăsta. Poate de data asta nu mai mușcă nimeni.",
"RoLlama2-7b-Instruct-Restyling-formal": "Dacă domnul Nicușor va deveni președinte, situația va fi una deosebit de dificilă pentru cetățeni. Multe promisiuni, dar puține realizări concrete. A mai înșelat așteptările oamenilor în calitate de primar, iar acum revine cu aceleași promisiuni electorale. Este momentul ca alegătorii să privească critic și rațional aceste declarații."
}
### Data Fields
- `id` - unique ID for each text instance
- `human` - human written text
- `RoLlama2-7b-Instruct-Completion` - text generated by RoLlama2-7b-Instruct model using a completion task
- `RoLlama2-7b-Instruct-Paraphrasing` - text generated by RoLlama2-7b-Instruct model using a paraphrasing task
- `RoLlama2-7b-Instruct-Restyling-informal` - text generated by RoLlama2-7b-Instruct model using a restyling (informal) task
- `RoLlama2-7b-Instruct-Restyling-formal` - text generated by RoLlama2-7b-Instruct model using a restyling (formal) task
### Data Splits
tbd
## Dataset Creation
### Curation Rationale
tbd
### Source Data
- Social networks such as TikTok, X and Reddit - public data (using crawling via API or legal scraping, including manual scraping)
- Political topics (presidential, local, parties, ideology, etc.)
### Annotations
- Manual annotations - toxicity score, slang presence, figurative language presence, HYBRID text labeling
- Semi-automatic - using LLMs + human validation (?)
- Consensus on at least 2 native annotators for validation (tbd)
tbc (annotation process)
### Personal and Sensitive Information
The data was public at the time of collection. PII removal has been performed.
## Considerations for Using the Data
### Social Impact of Dataset
tbd
### Discussion of Biases
tbd
### Other Known Limitations
tbd
### Ethical Considerations
- Comments are completely anonymized.
- Text generated by LLMs is clearly marked and is not presented as real content.
- No political ideas/ideologies are promoted — the purpose is synthetic text detection, linguistic and stylistic analysis.
- Use of this dataset outside of academic purposes is not encouraged.
## Additional Information
### Dataset Curators
tbd
### Licensing Information
All texts are collected respecting the platform's terms of service and consent policies. The dataset is intended for academic and non-commercial research.
tbc
### Citation Information
tbd
### Contributions
tbd