kona2-12B-Instruct/README.md

---
license: apache-2.0
language:
  - ka
  - en
  - multilingual
library_name: transformers
pipeline_tag: text-generation
tags:
  - llm
  - georgian
  - instruct
  - chat
  - function-calling
  - conversational
datasets:
  - tbilisi-ai-lab/kona-sft-mix-2.6M
  - tbilisi-ai-lab/kona-sft-function-calling-115k
  - tbilisi-ai-lab/kona-sft-function-calling-ka-93k
base_model:
  - tbilisi-ai-lab/kona2-12B-Base
---

# Kona2-12B-Instruct

**Kona2-12B-Instruct** is a 12-billion parameter instruction-tuned language model for Georgian and English. Built on [Kona2-12B-Base](https://huggingface.co/tbilisi-ai-lab/kona2-12B-Base) through supervised fine-tuning (SFT), it excels at chat, question answering, and function calling.

## Model Summary

| Property | Value |
|----------|-------|
| Parameters | 12B |
| Architecture | Mistral (Transformer) |
| Context Length | 32K tokens |
| Languages | Georgian (ka), English (en), other (limited) |
| Training | Supervised Fine-Tuning (SFT) |
| Training Examples | ~2.8M instructions |
| Function Calling | Yes (Hermes format) |
| Base Model | [kona2-12B-Base](https://huggingface.co/tbilisi-ai-lab/kona2-12B-Base) |

## Intended Uses

### Primary Use Cases
- Conversational AI assistants (Georgian/English)
- Question answering and information retrieval
- Function/tool calling applications
- **Translation between Georgian and English** (especially strong)
- Code generation and explanation
- Educational and tutoring applications

## Training

### Training Data

| Dataset | Examples | Description |
|---------|----------|-------------|
| [kona-sft-mix-2.6M](https://huggingface.co/datasets/tbilisi-ai-lab/kona-sft-mix-2.6M) | 2,606,173 | Mixed instruction dataset (KA/EN) |
| [kona-sft-function-calling-115k](https://huggingface.co/datasets/tbilisi-ai-lab/kona-sft-function-calling-115k) | ~115K | Function calling (English) |
| [kona-sft-function-calling-ka-93k](https://huggingface.co/datasets/tbilisi-ai-lab/kona-sft-function-calling-ka-93k) | ~93K | Function calling (Georgian) |

**Data Sources Include:**
- Wikipedia Q&A (RAFT-generated)
- Orca-style reasoning
- Self-instruct (Alpaca-style)
- Translation pairs (EN-KA)
- Code instructions
- Math instructions
- PersonaHub reasoning
- Glaive & Hermes function calling

### Training Procedure

- **Method:** Supervised Fine-Tuning (SFT)
- **LoRA Config:** r=256, alpha=512
- **Learning Rate:** 3e-5
- **Epochs:** 2
- **Training Context:** 32K tokens
- **Packing:** Enabled
- **Precision:** BF16
- **Infrastructure:** DeepSpeed ZeRO-2

## Usage

### Installation

```bash
pip install transformers torch accelerate
```

### Chat Completion

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

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

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "რა არის საქართველოს დედაქალაქი?"}
]

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

outputs = model.generate(inputs, max_new_tokens=256, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### Function Calling (Hermes Format)

See the tokenizer's jinja template (`tokenizer_config.json`) for details on how function calling is formatted.

```python
tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get current weather for a location",
            "parameters": {
                "type": "object",
                "properties": {
                    "location": {"type": "string", "description": "City name"}
                },
                "required": ["location"]
            }
        }
    }
]

messages = [
    {"role": "system", "content": "You are a helpful assistant with access to tools."},
    {"role": "user", "content": "What's the weather in Tbilisi?"}
]

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

outputs = model.generate(inputs, max_new_tokens=256)
# Output will include <tool_call>{"name": "get_weather", "arguments": {"location": "Tbilisi"}}</tool_call>
```

## Related Models

| Model | Description |
|-------|-------------|
| [kona2-12B-Base](https://huggingface.co/tbilisi-ai-lab/kona2-12B-Base) | Pre-trained base model |
| [kona2-12B](https://huggingface.co/tbilisi-ai-lab/kona2-12B) | DPO-aligned version (recommended) |
| [kona2-small-3.8B](https://huggingface.co/tbilisi-ai-lab/kona2-small-3.8B) | Smaller 3.8B model |

## Limitations

- Training data cutoff: 2024

## Technical Specifications

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

## Citation

```bibtex
@misc{tbilisi2025kona2instruct,
  title        = {Kona2-12B-Instruct: A Georgian Instruction-Tuned Language Model},
  author       = {Tbilisi AI Lab Team},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/tbilisi-ai-lab/kona2-12B-Instruct}}
}
```

## License

This model is released under the [Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0).

## Contact

- **Organization:** [Tbilisi AI Lab](https://huggingface.co/tbilisi-ai-lab)
- **Website:** [ailab.ge](https://ailab.ge)
- **Chat:** [chat.ailab.ge](https://chat.ailab.ge)
- **API:** [api.ailab.ge](https://api.ailab.ge)