Instructions to use kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru with Transformers:

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

pipe = pipeline("text-generation", model="kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru")
model = AutoModelForCausalLM.from_pretrained("kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

llama-cpp-python

How to use kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru",
	filename="gguf/minicpm5-1b-fable5-twix-ru-f16.gguf",
)

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

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16
# Run inference directly in the terminal:
llama cli -hf kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16
# Run inference directly in the terminal:
llama cli -hf kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16

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 kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16
# Run inference directly in the terminal:
./llama-cli -hf kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16

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 kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16

Use Docker

docker model run hf.co/kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16

LM Studio
Jan

vLLM

How to use kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16

SGLang

How to use kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru 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 "kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru with Ollama:
```
ollama run hf.co/kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16
```

Unsloth Studio

How to use kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru 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 kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru 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 kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru to start chatting

How to use kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16

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": "kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16

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 kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16

Run Hermes

hermes

Atomic Chat new

OpenClaw new

How to use kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Docker Model Runner
How to use kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru with Docker Model Runner:
```
docker model run hf.co/kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16
```

Lemonade

How to use kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull kaufkino/MiniCPM5-1B-Fable5-Thinking-T-Wix-ru:F16

Run and chat with the model

lemonade run user.MiniCPM5-1B-Fable5-Thinking-T-Wix-ru-F16

List all available models

lemonade list

kaufkino commited on 2 days ago

Commit

56324c6

verified ·

1 Parent(s): 588da1a

v2: reasoning-SFT model card

Browse files

Files changed (1) hide show

README.md +18 -16

README.md CHANGED Viewed

@@ -19,33 +19,34 @@ tags:
 # MiniCPM5-1B-Fable5-Thinking-T-Wix-ru
-> ⚠️ **Экспериментальный merge.** Не обучалась как единое целое: это LoRA-адаптер, обученный на другой модели, применённый к дистилляту. Известные дефекты описаны ниже. Метрики не снимались. Не для продакшена.
-Экспериментальное объединение двух моделей:
-- **[GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking](https://huggingface.co/GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking)** — дистиллят Claude Opus/Fable5 (thinking, tool calling) на базе MiniCPM5-1B
-- **T-Wix LoRA (r=32)** из [kaufkino/MiniCPM5-1B-T-Wix-ru](https://huggingface.co/kaufkino/MiniCPM5-1B-T-Wix-ru) — русификация на ~461k диалогов [t-tech/T-Wix](https://huggingface.co/datasets/t-tech/T-Wix)
-Метод: LoRA-адаптер, обученный на базовой `openbmb/MiniCPM5-1B`, применён к дистилляту через `PeftModel.from_pretrained` + `merge_and_unload()` (адаптер и дистиллят имеют общую базу и архитектуру `LlamaForCausalLM`).
 ## Наблюдаемое поведение (ручной смоук-тест, не бенчмарк)
-**Что перенеслось:**
-- Беглый русский с хорошей структурой (наследие T-Wix LoRA)
-- Простая математика с пошаговым решением — заметно лучше, чем у чистой T-Wix-версии (наследие дистиллята)
 - Tool calling: корректно выбирает функцию и аргументы
 **Известные дефекты:**
-- Склонность к зацикливанию — **обязательно используйте `repetition_penalty=1.1–1.3`**
-- `<think>`-режим дистиллята после merge в явном виде почти не проявляется
-- Фактология на русском слабая (модель 1B + языковой перенос): для фактов используйте RAG
 - Перевод EN→RU нестабилен
 ## Структура репозитория
 | Путь | Содержимое |
 |---|---|
-| `/` | merged-модель, bf16 safetensors |
 | `/gguf` | GGUF: F16, Q8_0 |
 ## Запуск
@@ -59,12 +60,13 @@ tokenizer = AutoTokenizer.from_pretrained(repo)
 messages = [{"role": "user", "content": "Реши: 17 * 23 = ?"}]
 inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True,
                                        return_tensors="pt", return_dict=True).to(model.device)
-out = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.95,
-                     do_sample=True, repetition_penalty=1.15)
 print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
 ```
-Chat template (включая tools-формат) унаследован от дистиллята.
 ## Инфраструктура
@@ -72,4 +74,4 @@ Chat template (включая tools-формат) унаследован от д
 ---
-**EN:** Experimental merge: T-Wix Russian LoRA (r=32, trained on base MiniCPM5-1B) applied onto a Claude Opus/Fable5 distill of the same base. Russian fluency + improved math/tool-calling vs the plain T-Wix tune; known issues: repetition loops (use repetition_penalty≥1.1), suppressed think-mode, weak Russian factual recall. Research artifact.

 # MiniCPM5-1B-Fable5-Thinking-T-Wix-ru
+> ⚠️ **Экспериментальная модель.** Метрики не снимались (только ручной смоук-квиз). Не для продакшена.
+Русскоязычная 1B-модель с работающим `<think>`-режимом. Получена в два этапа:
+1. **Merge**: T-Wix LoRA (r=32, обучена на базовой `openbmb/MiniCPM5-1B` на ~461k русских диалогов [t-tech/T-Wix](https://huggingface.co/datasets/t-tech/T-Wix)) применена к дистилляту [GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking](https://huggingface.co/GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking) через `PeftModel` + `merge_and_unload()`.
+2. **Reasoning-SFT (v2, текущая версия)**: merge дообучен LoRA (r=16, LR 2e-5, 1 эпоха) на reasoning-сабсете T-Wix — 31k примеров с русскими `<think>`-трейсами; управляющие суффиксы ` /think` из данных убраны, чтобы thinking стал поведением по умолчанию.
 ## Наблюдаемое поведение (ручной смоук-тест, не бенчмарк)
+**Что работает:**
+- `<think>` включается сам на задачах, где уместен (математика, рассуждения), думает на языке вопроса (RU-вопрос → русский трейс, EN → английский), закрывает блок и даёт ответ
+- Пошаговая математика с самопроверкой внутри think
 - Tool calling: корректно выбирает функцию и аргументы
+- Беглый русский с хорошей структурой
 **Известные дефекты:**
+- На длинных выкладках think может не уложиться в маленький `max_new_tokens` — ставьте ≥1000–2000
+- Фактология на русском слабая (модель 1B): для фактов используйте RAG
 - Перевод EN→RU нестабилен
+- Рекомендуется `repetition_penalty=1.1` (без него возможны циклы)
 ## Структура репозитория
 | Путь | Содержимое |
 |---|---|
+| `/` | модель после reasoning-SFT, bf16 safetensors |
 | `/gguf` | GGUF: F16, Q8_0 |
+| `/lora` | reasoning-LoRA (r=16) отдельно — применима к merge-версии из истории коммитов |
 ## Запуск
 messages = [{"role": "user", "content": "Реши: 17 * 23 = ?"}]
 inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True,
                                        return_tensors="pt", return_dict=True).to(model.device)
+out = model.generate(**inputs, max_new_tokens=1500, temperature=0.5, top_p=0.95,
+                     do_sample=True, repetition_penalty=1.1)
 print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
 ```
+Chat template (включая tools-формат и `enable_thinking`) унаследован от дистиллята.
+Принудительно выключить размышления: `enable_thinking=False` в `apply_chat_template`.
 ## Инфраструктура
 ---
+**EN:** Russian 1B model with working think-mode: T-Wix Russian LoRA merged onto a Claude Opus/Fable5 distill, then SFT'd on the T-Wix reasoning subset (31k Russian think-traces, r=16, 1 epoch). Think auto-triggers on reasoning tasks in the question's language and closes properly; tool calling preserved. Known issues: weak Russian factual recall (use RAG), unstable EN→RU translation, use repetition_penalty=1.1 and max_new_tokens≥1000. Research artifact, no formal evals.