tbilisi-ai-lab/kona2-12B-Base

Kona2-12B-Base

Kona2-12B-Base is a 12-billion parameter base language model optimized for Georgian language understanding and generation. It is built by continue-pretraining Mistral-Nemo-Base-2407 on approximately 30 billion tokens of Georgian/English data with expanded Georgian vocabulary.

Model Summary

Intended Uses

Primary Use Cases

• —Base model for Georgian language fine-tuning
• —Georgian text generation and completion
• —Multilingual text understanding (KA/EN primary, others limited)
• —Foundation for instruction-tuned models
• —Translation capabilities (enhanced in fine-tuned versions)

Training

Training Data (~30B Tokens)

• —Open source corpora
• —Web content (custom crawlers/scrapers)
• —Translated texts

Vocabulary Expansion

Added ~20K Georgian tokens to improve tokenization efficiency:

• —Tokenizer fertility: 1.9 tokens/word on Georgian text
• —New embeddings initialized as mean of subtoken embeddings

Training Procedure

• —Method: Continue pre-training
• —Embeddings: Full training (unfrozen)
• —LoRA: High-rank adaptation on transformer layers
• —Training Context: 8K tokens
• —Precision: BF16
• —Infrastructure: NVIDIA H100 GPUs

Usage

Installation

pip install transformers torch accelerate

Basic Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "tbilisi-ai-lab/kona2-12B-Base",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("tbilisi-ai-lab/kona2-12B-Base")

# Text completion
text = "საქართველო არის"
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Related Models

Limitations

• —Training data cutoff: 2024

Training Pipeline

Mistral-Nemo-Base-2407
    │
    ├── Expand Vocabulary (+20K Georgian tokens)
    │   └── Initialize with token average
    │
    └── Continue Pre-training (~30B tokens)
        └── Full embeddings training + high-rank LoRA
            │
            └── kona2-12B-Base ← YOU ARE HERE
                │
                └── SFT (~2.8M examples)
                    │
                    └── kona2-12B-Instruct
                        │
                        └── DPO (387K pairs)
                            │
                            └── kona2-12B

Technical Specifications

• —Precision: BF16/FP16 supported
• —Minimum VRAM: 24GB (with quantization)

Citation

@misc{tbilisi2025kona2base,
  title        = {Kona2-12B-Base: A Georgian Language Model},
  author       = {Tbilisi AI Lab Team},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/tbilisi-ai-lab/kona2-12B-Base}}
}

License

This model is released under the Apache 2.0 License.

Contact

• —Organization: Tbilisi AI Lab
• —Website: ailab.ge
• —API: api.ailab.ge