RooseBERT-large-scr-cased

This model is a fine-tuned version of bert-large-cased. It achieves the following results on the evaluation set:

• Loss: 0.9116
• Accuracy: 0.7799
• Perplexity 2.601

Model description

This model builds on the same architecture as bert-base-cased, leveraging transformer-based contextual embeddings to better understand the nuances of political language.

Intended Use Cases

Suitable Applications

• Political discourse analysis: Identifying patterns, sentiments, and rhetoric in debates.
• Contextual word interpretation: Understanding the meaning of words within political contexts.
• Sentiment classification: Differentiating positive, neutral, and negative sentiments in political speech.
• Text generation improvement: Enhancing auto-completions and summaries in politically focused language models.

Limitations

• Bias Sensitivity: Since it was trained on political debates, inherent biases in the data may be reflected in the model’s outputs.
• Not Suitable for General-Purpose NLP: Its optimization is specific for political contexts.
• Does Not Perform Fact-Checking: The model does not verify factual accuracy.

Training and Evaluation Data

The model was trained on a curated dataset of political debates sourced from:

• Parliamentary transcripts
• Presidential debates and public speeches

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 64
• eval_batch_size: 64
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 4
• total_train_batch_size: 2048
• total_eval_batch_size: 512
• optimizer: Use adamw_torch with betas=(0.9,0.98) and epsilon=1e-06 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• training_steps: 125000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Accuracy	Validation Loss
No log	0	0	0.0000	10.5234
1.2678	12.6967	50000	0.7217	1.2314
1.121	25.3936	100000	0.7453	1.0977
0.9192	137.3328	125000	0.9111	0.7799

Framework versions

• Transformers 4.49.0.dev0
• Pytorch 2.5.1
• Datasets 3.2.0
• Tokenizers 0.21.0

Citation

If you use this model, cite us:

@misc{
    dore2025roosebertnewdealpolitical, 
    title={RooseBERT: A New Deal For Political Language Modelling}, 
    author={Deborah Dore and Elena Cabrio and Serena Villata}, 
    year={2025}, 
    eprint={2508.03250}, 
    archivePrefix={arXiv}, 
    primaryClass={cs.CL}, 
    url={https://arxiv.org/abs/2508.03250}, 
}