Model Card for PoliWings/opposing-views-left-wing

This model is a fine-tuned version of Bielik-11B-v2.2-Instruct, designed to simulate a left-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 left-wing ideological perspectives in the Polish language. It can be utilized to generate political statements and interpret public discourse from a left-leaning point of view.

Bias, Risks, and Limitations

This model explicitly contains a strong political bias by design.

Experiments show that fine-tuning the base model on left-wing speeches amplifies its preexisting left-leaning bias.
System prompts can have a stronger influence on ideological responses than the fine-tuning itself.
Strong role instructions can largely override the fine-tuning effects.

Recommendations

Users should be aware that the model is not a neutral observer and explicitly reflects left-wing ideological tendencies. 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 Left-wing subset of the newly constructed Polish Parliamentary Speeches Corpus (PPSC).

The subset contains 10,594 speeches delivered by 51 left-wing politicians.
The data spans from November 2011 to April 2025.
The average speech length in this subset is 243.30 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%
Left-wing fine-tuned	82.85% (+10.43 pp)

Under a basic prompt, the left-wing model achieved an aggregated left-bias score of 82.85%. The model exhibited its strongest left-leaning stance in Climate Policy, reaching up to 96% under explicit left prompts.