uaebn/leesplank-municipal

Leesplank Municipal Finetune

A specialized adaptation of UWV/leesplank-noot-eurollm-1.7b for simplifying Dutch municipal and legal texts to B1 reading level.

Model Description

This model is fine-tuned specifically for the municipal domain, building upon the UWV Leesplank Noot base model. The fine-tuning focused on adapting the model's behavior to handle the specific vocabulary, structures, and simplification patterns common in Dutch municipal communications.

Intended Use

• Simplifying Dutch municipal documents and legal texts
• Making government communications more accessible to B1 readers
• Supporting clear language initiatives in public administration

Performance

The model has been evaluated on a human-verified municipal text dataset:

Metric	Score	Description
SARI	56.18	Measures simplification quality (higher is typically better*)
BERTScore	0.95	Semantic meaning preservation
LiNT-II	48.76 - 2.69	Dutch readability score (B1 target)
Flesch-Douma	52.68	Reading ease score

*Depends on simplification strategy.

Evaluation dataset: temp

Intended use

Use the ChatML format, and prefix your input text with "Vereenvoudig:" only.

message = [
    {
        "role": "user", 
        "content": f"Vereenvoudig: {input_text}"
    }
]

Training Details

Base Model: UWV/Leesplank-Noot-1.7B

Method: QLoRA fine-tuning with targeted parameter updates to specialize for municipal domain

Configuration:

• LoRA Rank: 64
• LoRA Alpha: 128
• Target Modules: "q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"

Limitations

• Optimized specifically for Dutch municipal and legal texts
• Performance may vary on other text types
• Inherits base model limitations

Acknowledgments

This model builds upon the work by UWV on the Leesplank Noot base model.