Mecellem - Turkish Legal Language Models

Unified repository for Turkish legal language model benchmarks and training pipelines.

Model Collection: https://huggingface.co/collections/newmindai/mecellem-models

📁 Project Structure

Mecellem/
├── benchmark/
│   ├── mlm/                    # MLM (Masked Language Modeling) benchmark
│   └── embedding_model/        # Embedding model evaluation benchmark
├── training/
│   └── post-training-retrieval/ # Post-training retrieval model training
└── README.md

🎯 Components

Benchmark - MLM

Masked Language Modeling benchmark for Turkish BERT models.

Location: benchmark/mlm/

Features:

• BERT's original 80-10-10 masking strategy
• Multiple masking ratios (5%, 10%, 15%)
• Multi-dataset evaluation (QA, Review, Biomedical, Legal)
• ModernBERT architecture support

Quick Start:

cd benchmark/mlm
python benchmark_mlm.py

See benchmark/mlm/README.md for detailed documentation.

Benchmark - Embedding Model

Comprehensive evaluation framework for Turkish embedding models using MTEB.

Location: benchmark/embedding_model/

Features:

• MTEB integration for Turkish tasks
• Legal domain task evaluation
• Model2Vec adapter support
• Automatic max_seq_length optimization

Quick Start:

cd benchmark/embedding_model
./run_benchmark.sh

See benchmark/embedding_model/README.md for detailed documentation.

Training - Post-Training Retrieval

Post-training pipeline for retrieval-optimized Turkish legal embedding models.

Location: training/post-training-retrieval/

Features:

• Sentence Transformers integration
• MSMARCO-TR dataset support
• Multiple loss functions (CachedGIST, MultipleNegativesRanking, etc.)
• Unified configuration system

Quick Start:

cd training/post-training-retrieval
./launchers/train.sh mursit_base_tr_retrieval

See training/post-training-retrieval/README.md for detailed documentation.

🚀 Getting Started

Prerequisites

• Python 3.8+
• CUDA-capable GPU (recommended)
• Git

Installation

Each component has its own requirements. Please refer to the individual README files for specific installation instructions.

📚 Documentation

• MLM Benchmark Documentation
• Embedding Model Benchmark Documentation
• Post-Training Retrieval Documentation

🤝 Contributing

This repository consolidates three separate projects:

• MLM benchmark (benchmark/mlm)
• Embedding model evaluation (benchmark/embedding_model)
• Post-training retrieval (training/post-training-retrieval)

All code and README files have been preserved from the original projects.

📄 License

This project is licensed under the Apache License 2.0. See the LICENSE file for details.

📖 Citation

If you use Mecellem models in your research, please cite our paper:

@article{mecellem2026,
  title={Mecellem Models: Turkish Models Trained from Scratch and Continually Pre-trained for the Legal Domain},
  author={Uğur, Özgür and Göksu, Mahmut and Çimen, Mahmut and Yılmaz, Musa and Şavirdi, Esra and Demir, Alp Talha and Güllüce, Rumeysa and Çetin, İclal and Sağbaş, Ömer Can},
  journal={arXiv preprint arXiv:2601.16018},
  year={2026},
  month={January},
  url={https://arxiv.org/abs/2601.16018},
  doi={10.48550/arXiv.2601.16018},
  eprint={2601.16018},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}

📧 Contact

For questions or issues, please contact: aiteam@newmind.ai