---
base_model: Qwen/Qwen3-14B-Base
license: apache-2.0
language:
  - uz
library_name: transformers
pipeline_tag: text-generation
tags:
  - text-generation
  - qwen3
  - uzbek
  - cyrillic
  - lora
  - unsloth
---

# Qwen3-14B-Base-Uzbek-Cyrillic

A fine-tuned version of [Qwen/Qwen3-14B-Base](https://huggingface.co/Qwen/Qwen3-14B-Base)
adapted for the Uzbek language in Cyrillic script. The model was trained with
LoRA using the [Unsloth](https://github.com/unslothai/unsloth) framework, which
improves fluency and grammatical coherence on Uzbek (Cyrillic) text while
retaining the multilingual capabilities of the base model.

## Model Details

| Property | Value |
| --- | --- |
| Base model | Qwen/Qwen3-14B-Base |
| Architecture | Transformer decoder (causal LM) |
| Parameters | 14.8B |
| Context length | 32,768 tokens |
| Fine-tuning method | LoRA (r=16, alpha=32, dropout=0.0) |
| Training framework | Unsloth |
| Precision | bfloat16 |
| Language | Uzbek (Cyrillic), multilingual |
| License | Apache 2.0 |

This is a base (non-instruction-tuned) model. It is intended for text
completion and continued pretraining workflows rather than turn-based chat
out of the box. For conversational use, apply your own chat template or further
instruction fine-tuning.

## Intended Use

- Text generation and completion in Uzbek (Cyrillic)
- Summarization and content generation in Uzbek
- Multilingual applications targeting Central Asian languages
- A starting point for further task- or instruction-specific fine-tuning

## Limitations

- The model is not instruction-tuned and may not follow prompts as a chat model would.
- Output may contain factual errors or biases inherited from the base model and training data.
- Performance on Latin-script Uzbek or other scripts is not the focus of this fine-tune.
- Generated text should be reviewed before use in production or sensitive contexts.

## Usage

### Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Just-Bax/Qwen3-14B-Base-Uzbek-Cyrillic"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

prompt = "Ассалому алайкум! Бугунги об-ҳаво ҳақида маълумот:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### Pipeline

```python
from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="Just-Bax/Qwen3-14B-Base-Uzbek-Cyrillic",
    torch_dtype="bfloat16",
    device_map="auto",
)
print(pipe("Ўзбекистон пойтахти", max_new_tokens=64)[0]["generated_text"])
```

### vLLM

```bash
pip install vllm
vllm serve "Just-Bax/Qwen3-14B-Base-Uzbek-Cyrillic"
```

```bash
curl -X POST "http://localhost:8000/v1/completions" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "Just-Bax/Qwen3-14B-Base-Uzbek-Cyrillic",
    "prompt": "Бир бор экан, бир йўқ экан,",
    "max_tokens": 512,
    "temperature": 0.5
  }'
```

### Unsloth

```python
from unsloth import FastModel

model, tokenizer = FastModel.from_pretrained(
    model_name="Just-Bax/Qwen3-14B-Base-Uzbek-Cyrillic",
    max_seq_length=2048,
)
```

## Training

The model was fine-tuned with LoRA adapters (r=16, alpha=32, dropout=0.0) on
Uzbek Cyrillic text using Unsloth in bfloat16 precision. The adapters were
merged into the base weights for distribution, so the model can be loaded
directly with `transformers` without additional adapter loading.

## License

Released under the Apache 2.0 license, consistent with the base model
[Qwen/Qwen3-14B-Base](https://huggingface.co/Qwen/Qwen3-14B-Base).

## Citation

If you use this model, please cite the base model and the Unsloth framework:

```bibtex
@misc{qwen3,
  title  = {Qwen3},
  author = {Qwen Team},
  year   = {2025},
  url    = {https://huggingface.co/Qwen/Qwen3-14B-Base}
}
```