Llama-3.1-8B-madon-arg-classification

This model is a fine-tuned version of Llama-3.1-8B for multi-label classification of legal arguments in Czech court decisions. It was introduced in the paper Mining Legal Arguments to Study Judicial Formalism.

Model Description

The model is part of the MADON project, specifically focusing on Task 2: Argument type classification. It is trained to classify paragraphs from decisions of the Czech Supreme Courts into eight traditional types of legal arguments:

• LIN: Linguistic
• SI: Systemic
• CL: Case-law
• D: Doctrinal
• HI: Historical
• PL: Teleological (Purposive)
• TI: Logical
• PC: Principles/Comparative

The model was adapted to the Czech legal domain through continued pretraining on a corpus of over 300,000 court decisions and fine-tuned using asymmetric loss and class weighting to address dataset imbalance.

• Paper: Mining Legal Arguments to Study Judicial Formalism
• Repository: trusthlt/madon
• Language: Czech (cs)

Usage

If you want to use it for multi-label classification of Czech legal arguments, we suggest:

from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline

model = AutoModelForSequenceClassification.from_pretrained("TrustHLT/Llama-3.1-8B-madon-arg-classification")
tokenizer = AutoTokenizer.from_pretrained("TrustHLT/Llama-3.1-8B-madon-arg-classification")

pipe = pipeline("text-classification", model=model, tokenizer=tokenizer, return_all_scores=True, function_to_apply="none")

text = "This is a legal paragraph"

print(pipe(text))

Citation

If you use this model or the MADON dataset in your research, please cite:

@article{madon2025mining,
  title={Mining Legal Arguments to Study Judicial Formalism},
  author={Anonymous},
  journal={arXiv preprint arXiv:2512.11374},
  year={2025}
}