Instructions to use stukenov/sozkz-mgpt-1.3b-translate-kkru-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use stukenov/sozkz-mgpt-1.3b-translate-kkru-v1 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="stukenov/sozkz-mgpt-1.3b-translate-kkru-v1")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("stukenov/sozkz-mgpt-1.3b-translate-kkru-v1")
model = AutoModelForCausalLM.from_pretrained("stukenov/sozkz-mgpt-1.3b-translate-kkru-v1", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use stukenov/sozkz-mgpt-1.3b-translate-kkru-v1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "stukenov/sozkz-mgpt-1.3b-translate-kkru-v1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "stukenov/sozkz-mgpt-1.3b-translate-kkru-v1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/stukenov/sozkz-mgpt-1.3b-translate-kkru-v1

SGLang

How to use stukenov/sozkz-mgpt-1.3b-translate-kkru-v1 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "stukenov/sozkz-mgpt-1.3b-translate-kkru-v1" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "stukenov/sozkz-mgpt-1.3b-translate-kkru-v1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "stukenov/sozkz-mgpt-1.3b-translate-kkru-v1" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "stukenov/sozkz-mgpt-1.3b-translate-kkru-v1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use stukenov/sozkz-mgpt-1.3b-translate-kkru-v1 with Docker Model Runner:
```
docker model run hf.co/stukenov/sozkz-mgpt-1.3b-translate-kkru-v1
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

SozKZ mGPT 1.3B Translate KK-RU v1

A bidirectional Kazakh-Russian translation model based on mGPT-1.3B-kazakh. Trained in two stages on 500K parallel sentence pairs from EkiTil Parallel Corpus.

Model Details


Base model	ai-forever/mGPT-1.3B-kazakh
Architecture	GPT-2 (24 layers, 2048 hidden, 16 heads)
Parameters	1.42B
Languages	Kazakh (kk) <-> Russian (ru)
License	MIT
Training data	stukenov/ekitil-parallel-kkru-v2 (500K pairs)
Hardware	1x NVIDIA H100 80GB SXM
Total training time	~6.5 hours

Training Pipeline

Stage 1: Continual Pretraining (full fine-tune)

Format: [KK>RU] source [SEP] target</s> and [RU>KK] source [SEP] target</s>
1M examples (500K pairs x 2 directions), 1 epoch
BS=32, grad_accum=2, lr=2e-5, cosine schedule
Eval loss: 1.054

Stage 2: SFT with LoRA (instruction format)

Format: ### Аудар [KK>RU]:\nsource\n### Аударма:\ntarget</s>
LoRA r=32, alpha=64, targets: c_attn, c_proj, c_fc (25M trainable params)
1M examples, 1 epoch
Eval loss: 0.896

Prompt Format

Kazakh to Russian

### Аудар [KK>RU]:
Қазақстан Республикасы — Орталық Азиядағы мемлекет.
### Аударма:

Russian to Kazakh

### Аудар [RU>KK]:
Казахстан — государство в Центральной Азии.
### Аударма:

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "stukenov/sozkz-mgpt-1.3b-translate-kkru-v1",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("stukenov/sozkz-mgpt-1.3b-translate-kkru-v1")

# Kazakh -> Russian
prompt = "### Аудар [KK>RU]:\nҚазақстан Республикасы — Орталық Азиядағы мемлекет.\n### Аударма:\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=100, do_sample=False, repetition_penalty=1.2)

print(tokenizer.decode(out[0], skip_special_tokens=True))
# -> Республика Казахстан - государство в Центральной Азии.

Translation Examples

Direction	Input	Output
KK->RU	Қазақстан Республикасы — Орталық Азиядағы мемлекет.	Республика Казахстан - государство в Центральной Азии.
RU->KK	Казахстан — государство в Центральной Азии.	Қазақстанда Орталық Азияда мемлекет орналасады.
KK->RU	Абай Құнанбайұлы — ұлы қазақ ақыны, ағартушы, ойшыл.	Академик Абая Кунанбаева - выдающийся казахский поэт и просветитель.
RU->KK	Алматы — крупнейший город Казахстана и культурная столица.	Астана Қазақстанның елордасы және мәдени орталығы болып табылады.
RU->KK	Образование является важнейшим направлением политики.	Білім мемлекеттік саясаттың мақсаты болып табылады.

Limitations

Repetition issues on longer outputs (needs repetition_penalty=1.2+)
Some factual errors in translation (e.g., names may be altered)
Decoder-only architecture is less optimal for translation than encoder-decoder (T5/NLLB)
Not evaluated on FLORES+ benchmark yet

Related Models

ai-forever/mGPT-1.3B-kazakh - base model
stukenov/sozkz-mgpt-1.3b-kk-instruct-v1 - instruction-following variant
deepvk/kazRush-kk-ru - T5-based kk->ru translation (197M)

Citation

@misc{sozkz-mgpt-translate-2026,
  title={SozKZ mGPT 1.3B Translate KK-RU v1},
  author={Stukenov, Saken},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/stukenov/sozkz-mgpt-1.3b-translate-kkru-v1}
}

Downloads last month: -

Safetensors

Model size

1B params

Tensor type

BF16

Model tree for stukenov/sozkz-mgpt-1.3b-translate-kkru-v1

Base model

ai-forever/mGPT-1.3B-kazakh

Finetuned

(2)

this model

Dataset used to train stukenov/sozkz-mgpt-1.3b-translate-kkru-v1

Evaluation results

stage1_eval_loss
self-reported

1.054
stage2_eval_loss
self-reported

0.896