This model is a fine-tuned version of Qwen3-8B for conditional text generation of Slovak court decisions. Given the legal reasoning of a court case as input, the model generates the corresponding verdict/statement (výrok) of the decision.

Developed by: Bianka Šimková
Model type: Causal LLM fine-tuned for text generation
Language: Slovak
License: Apache license 2.0
Base model: Qwen/Qwen3-8B
Fine-tuned for: Diploma thesis — Using large language models for analyzing decisions of Slovak courts

Intended Use

Primary Use

Generating the verdict/statement (výrok) of a Slovak court decision conditioned on the legal reasoning (odôvodnenie) of the case.

Out-of-Scope Use

Court decisions in languages other than Slovak
Legal advice or real-world legal decision support — this model is a research prototype and should not be used for any actual legal proceedings
Use as an authoritative source of legal rulings

Training Data

The training data consists of Slovak court decisions sourced from open government records of the Slovak Republic. Each training example pairs the reasoning section of a decision as input with the verdict section as the generation target.

Data source: Open data portal of the Slovak Republic
Dataset format: JSON
Preprocessing: Text cleaning, segmentation into logical parts (reasoning / statement), sampling of examples under 2048 tokens

Training Procedure

Base model: Qwen3-8B
Fine-tuning method: LoRA
Framework: Hugging Face Transformers
Hardware: NVIDIA RTX A6000
Epochs: 1
Batch size: 1
Learning rate: 2e-4
Input format: Legal reasoning (odôvodnenie) of the court decision
Output format: Verdict/statement (výrok) of the court decision

Evaluation Results

Metric	Value
ROUGE-1	0.70
ROUGE-2	0.56
ROUGE-L	0.64
BLEU	0.40
Perplexity	1.89

Example Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("biankasimkova/qwen3-8b-finetuned-court-decisions-sk")
model = AutoModelForCausalLM.from_pretrained("biankasimkova/qwen3-8b-finetuned-court-decisions-sk")

reasoning = "Súd zistil, že žalobca ..."  # legal reasoning text
inputs = tokenizer(reasoning, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Limitations

Trained on Slovak court decisions only; performance on other legal systems or languages is not guaranteed
Generated verdicts may not reflect legally valid or current Slovak law — the model reproduces statistical patterns from training data
Not intended for production legal use

Citation

If you use this model, please cite the associated thesis:

@mastersthesis{Simkova2026,
  author    = {Bianka Šimková},
  title     = {Using large language models for analyzing decisions of Slovak courts},
  school    = {Institute of Artificial Intelligence,
Faculty of Electrical Engineering and Informatics, 
Technical University of Košice},
  year      = {2026}
}