Kona2-12B

Kona2-12B is the flagship 12-billion parameter Georgian language model from Tbilisi AI Lab. Built on Kona2-12B-Instruct and aligned using Direct Preference Optimization (DPO), it delivers higher quality, more helpful, and better-aligned responses.

This is the recommended model for production use.

Model Summary

Property	Value
Parameters	12B
Architecture	Mistral (Transformer)
Context Length	32K tokens
Languages	Georgian (ka), English (en), other (limited)
Training	Direct Preference Optimization (DPO)
Preference Pairs	387,319
Function Calling	Yes (Hermes format)
Base Model	kona2-12B-Instruct

Model Hierarchy

mistralai/Mistral-Nemo-Base-2407
    │
    ├── Expand Vocabulary (+20K Georgian tokens)
    │
    └── kona2-12B-Base (continue pre-training, ~30B tokens)
        │
        └── kona2-12B-Instruct (SFT on ~2.8M instructions)
            │
            └── kona2-12B (DPO on 387K preference pairs) ← YOU ARE HERE

Intended Uses

Primary Use Cases

• Production conversational AI (Georgian/English)
• High-quality question answering
• Function/tool calling with improved reliability
• Translation (especially strong)
• Content generation with better alignment
• Customer support automation

Training

DPO Training Data

Dataset	Pairs	Description
kona-dpo-mix-387k	387,319	Mixed preference pairs

DPO Pair Sources:

Source	Pairs	Strategy
Instruction Augmentation (Sonnet 3.5)	~29K	Claude 3.5 Sonnet as "chosen", base model as "rejected"
Function Calling (tools+call)	~5K	Correct tool calls vs text responses
Function Calling (tools, no call)	~5K	Correct text vs incorrect tool calls
Function Calling (no tools)	~5K	Appropriate responses without tools
Hermes Chats	~17K	High-quality conversation pairs
General Instructions	~320K	Mixed instruction following

DPO Scenarios

The model was trained on 4 distinct function-calling scenarios:

Scenario	Tools Available	Should Call	Chosen	Rejected
system_tools_yes_call_yes	Yes	Yes	Tool call	Text response
system_tools_yes_call_no	Yes	No	Text response	Tool call
tools_yes_call_yes	Yes	Yes	Tool call	Text response
tools_yes_call_no	Yes	No	Text response	Tool call

Training Procedure

• Method: Direct Preference Optimization (DPO)
• DPO Beta: 0.1
• LoRA Config: r=256, alpha=512
• Learning Rate: 5e-6
• Epochs: 2
• Training Context: 32K tokens
• Precision: BF16
• Infrastructure: DeepSpeed ZeRO-2

Usage

Installation

pip install transformers torch accelerate

Chat Completion

from transformers import AutoModelForCausalLM, AutoTokenizer

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

messages = [
    {"role": "system", "content": "You are a helpful, harmless, and honest assistant."},
    {"role": "user", "content": "დამეხმარე პითონზე ფუნქციის დაწერაში, რომელიც ითვლის ფაქტორიალს."}
]

inputs = tokenizer.apply_chat_template(
    messages, 
    return_tensors="pt",
    add_generation_prompt=True
).to(model.device)

outputs = model.generate(
    inputs, 
    max_new_tokens=512, 
    temperature=0.7,
    do_sample=True
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Related Models

Model	Stage	Use Case
kona2-12B-Base	Pre-training	Further fine-tuning
kona2-12B-Instruct	SFT	If DPO behavior is not desired
kona2-small-3.8B	All	Resource-constrained environments

Limitations

• Training data cutoff: 2024

Technical Specifications

• Precision: BF16/FP16 supported
• Minimum VRAM: 24GB (with 4-bit quantization)
• Recommended: 48GB+ for full precision

Citation

@misc{tbilisi2025kona2,
  title        = {Kona2-12B: A DPO-Aligned Georgian Language Model},
  author       = {Tbilisi AI Lab Team},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/tbilisi-ai-lab/kona2-12B}}
}

License

This model is released under the Apache 2.0 License.

Contact

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