parliabench-unsloth-gemma-2-9b

Model Name: <ParliaBench - unsloth/gemma-2-9b-bnb-4bit>

1. Model Summary

This is a fine-tuned model developed for the research paper "ParliaBench: An Evaluation and Benchmarking Framework for LLM-Generated Parliamentary Speech".

The model is trained to:

• Generate UK-style parliamentary debates
• Produce party-conditioned political arguments
• Model parliamentary dialogue structures
• Perform topic-aware generation using 21 EuroVoc topic categories

2. Base Model

Base model: "google/gemma-2-9b"

This repository contains only LoRA adapter weights.
The base model must be loaded separately.

3. Training Data

The model was trained using:

• ParlaMint-GB from Clarin
• Preprocessed into:
  • Structured debates
  • Speaker metadata
  • Section headers
  • 21-class topic labels

Dataset links:

• https://huggingface.co/datasets/argyrotsipi/train-dataset
• https://huggingface.co/datasets/argyrotsipi/generated-dataset

4. Training Procedure

Hyperparameters

• Batch Size 64
• Learning Rate 2e-4
• Max Steps 11194 (2 epochs)
• Warmup Steps 336 (10% of max steps for stability)
• Optimizer adamw
• Weight Decay 0.01
• Max Sequence Length 1024
• Scheduler linear

Hardware

GPU A100 AWS resources were provided by the National Infrastructures for Research and Technology GRNET and funded by the EU Recovery and Resiliency Facility.

• LoRA Rank (r) 16
• LoRA Alpha 16
• Target Modules 7 layers
• LoRA Dropout 0
• Bias Configuration none
• Random State 3407

5. Intended Uses

This model is designed for:

• Political debate simulation
• Parliamentary dialogue generation
• Academic NLP research
• Social behavior modeling

6. Limitations

• May reproduce political biases present in the corpus
• Not suitable for real political advice or predictions
• Model outputs are synthetic and not factual

7. Ethical Considerations

• Avoid using for political persuasion
• Use strictly in academic and research contexts
• Outputs may reflect UK political biases from the dataset

8. Citation

@misc{ParliaBench2025, title={ParliaBench: An Evaluation and Benchmarking Framework for LLM-Generated Parliamentary Speech}, author={Marios Koniaris and Argyro Tsipi and Panayiotis Tsanakas}, year={2025}, eprint={2511.08247}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2511.08247}, }

9. Authors

Marios Koniaris, Argyro Tsipi, Panayiotis Tsanakas ParliaBench: An Evaluation and Benchmarking Framework for LLM-Generated Parliamentary Speech.