Instructions to use RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf",
	filename="FP16.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M

Use Docker

docker model run hf.co/RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M

LM Studio
Jan

vLLM

How to use RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M

Ollama
How to use RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf with Ollama:
```
ollama run hf.co/RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M
```

Unsloth Studio new

How to use RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf to start chatting

Pi new

How to use RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M

Run Hermes

hermes

Docker Model Runner
How to use RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf with Docker Model Runner:
```
docker model run hf.co/RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M
```

Lemonade

How to use RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull RefalMachine/ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf:Q4_K_M

Run and chat with the model

lemonade run user.ruadapt_qwen2.5_3B_ext_u48_instruct_v4_gguf-Q4_K_M

List all available models

lemonade list

Описание модели

GGUF версия!! Инструктивная версия адаптированного на русский язык Qwen2.5-3B (RefalMachine/ruadapt_qwen2.5_3B_ext_u48_full_lr5e4_peft_mlp_32_32_bs256). В модели был заменен токенизатор, затем произведено дообучение (Continued pretraining) на русскоязычном корпусе, после чего была применена техника LEP (Learned Embedding Propagation, paper will be soon).

Благодаря новому токенизатору (расширенный tiktoken cl100k с помощью униграм токенизатора на 48 т. токенов) скорость генерации* русскоязычных текстов возрасла до 60% по сравнению с исходной моделью Qwen-2.5-3B-Instruct.

*Под скоростью генерации подразумевается количество русскоязычных символов/слов в секунду на одинаковых текстовых последовательностях.

Токенизация

Метрики и оценка качества

Модель была оценена на Ru-Arena-General, MERA, llmtf_open

Результаты на Ru-Arena-General

Замеры были произведены с использованием оффициального кода лидерборда (https://github.com/VikhrModels/ru_llm_arena), но с repetition_penalty=1.1.

Приведена лишь часть лидерборда, подробнее смотрите в репозитории бенчмарка (https://huggingface.co/spaces/Vikhrmodels/arenahardlb).

Model Name	Winrate	95% CI	Average # Tokens
gpt-4-1106-preview	90.9	( +1.3 / -0.9)	541
vikhr-nemo-12b-instruct-r-21-09-24	87.3	(+1.1 / -1.2)	627
gpt-4o-mini	83.9	(+1.9 / -1.6)	448
ruadapt_qwen2.5_7B_ext_u48_instruct	81.9	(+1.7 / -1.6)	556
gemma-2-9b-it	76.5	(+1.1 / -1.1)	459
Qwen2.5-7B-Instruct	76.0	(+1.6 / -1.8)	484
gemma-2-9b-it-sppo-iter3	73.6	(+2.1 / -2.2)	509
saiga_llama3_8b_v7	67.6	(+1.7 / -1.4)	503
ruadapt_qwen2.5_3B_ext_u48_instruct_v4	66.1	(+2.2 / -1.9)	531
t-lite-instruct-0.1	64.7	(+2.3 / -2.2)	810

Результаты на MERA

TODO

Результаты на llmtf_open

TODO

How to cite:

Tikhomirov M., Chernyshev D. Facilitating large language model Russian adaptation with Learned Embedding Propagation // 2024 (Preprint: https://arxiv.org/abs/2412.21140)

Tikhomirov M., Chernyshev D. Impact of Tokenization on LLaMa Russian Adaptation //2023 Ivannikov Ispras Open Conference (ISPRAS). – IEEE, 2023. – С. 163-168.

Результаты на MERA

Результаты на llmtf_open

TODO

Предупреждение

Ответы модели не отражают мнения авторов, а лишь повторяют знания полученные из данных на всех этапах обучения (предобучение, смена токенизатора, обучение на инструкциях, калибровка качества ответов). Модель была получена из сторонней предобученной модели, контроль за предобучением которой не является ответственностью текущих авторов. При создании данной версии модели не производилось никаких дополнительных действий, направленных на изменение заложенных в LLM "мнений". Используйте с осторожностью.