Model Card for PoliWings/opposing-views-right-wing
This model is a fine-tuned version of Bielik-11B-v2.2-Instruct, designed to simulate a right-leaning ideological perspective within the Polish political domain.
Model Details
Model Description
The model was developed to investigate political bias in Polish Large Language Models and explore the concept of agonistic pluralism by steering ideological tendencies. It has been fine-tuned on real political speeches from the Polish parliament.
- Developed by: Damian Jankowski, Radosław Gajewski, Jan Barczewski, Maciej Sikora, Jan Majkutewicz (Gdańsk University of Technology).
- Funded by: Supported by the "Cloud Artificial Intelligence Service Engineering (CAISE) platform" project (No. KPOD.05.10-IW.10-0005/24) as part of the European IPCEI-CIS program, financed by NRRP funds.
- Model type: Large Language Model (Fine-tuned using LoRA).
- Language: Polish.
- License: Apache 2.0
- Finetuned from model:
speakleash/Bielik-11B-v2.2-Instruct.
Model Sources
- Repository: https://github.com/PoliWings/SejmAI
- Dataset: Polish-Parliamentary-Speeches-Corpus
- Paper: "Asymmetric Ideological LLM Fine-Tuning in the Polish Political Domain"
Uses
Direct Use
The model is intended to be used for simulating distinct right-wing ideological perspectives in the Polish language. It can be utilized to generate political statements and interpret public discourse from a right-leaning point of view.
Bias, Risks, and Limitations
This model was specifically trained to adopt a right-wing perspective, however, it exhibits an asymmetric response to fine-tuning.
- Because the base model has a strong initial left-leaning bias, right-wing fine-tuning mainly counteracts this bias, often producing neutral responses rather than strongly right-leaning ones.
- Under basic prompts, the right-wing model remained slightly left-leaning.
- System prompts can have a stronger influence on ideological responses than the model's fine-tuned bias.
Recommendations
Users should understand that while the model is trained on right-wing data, it may still exhibit residual left-leaning tendencies on certain topics (like Climate Policy) due to the base model's foundation. The generated statements closely replicate real debate rhetoric and can be difficult to distinguish from authentic parliamentary speech.
Training Details
Training Data
The model was fine-tuned on the Right-wing subset of the newly constructed Polish Parliamentary Speeches Corpus (PPSC).
- The subset contains 9,966 speeches delivered by 56 right-wing politicians.
- The data spans from November 2011 to April 2025.
- The average speech length in this subset is 264.78 words.
Training Procedure
The dataset was processed by converting transcripts into instruction-response pairs using a synthetic instruction ("Speak on the following topic:").
Training Hyperparameters
- Training regime: The model was trained using LoRA applied to all linear layers.
- LoRA Rank: 32.
- LoRA $\alpha$: 32.
- Dropout: 0.1.
- Batch size: 2 per device.
- Learning rate: 1e-4 with cosine decay.
- Warm-up ratio: 0.1.
- Max sequence length: 2048.
Speeds, Sizes, Times
- Hardware: Trained on two NVIDIA H100 GPUs.
Evaluation
Testing Data, Factors & Metrics
- Testing Data: Evaluated using a custom "Political Views" benchmark consisting of 267 policy statements across five domains (economy, customary issues, foreign policy, political system, and climate policy).
- Metrics: An aggregated bias score was calculated as a percentage ratio, summing polarization weights scaled by stance strength. A score of 0 represents a perfectly neutral model.
Results
| Model | Left Bias Score |
|---|
| Base model (Bielik-11B-v2.2-Instruct) | 72.42% |
| Right-wing fine-tuned | 54.18% (-18.24 pp) |
Under a basic prompt, the right-wing fine-tuning reduced the base model's bias significantly to 54.18%. The model was most effective in adopting a right-leaning/neutral stance in Customary and Foreign Policy domains.
Framework versions
