Instructions to use nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl with Transformers:

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

pipe = pipeline("text-generation", model="nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl")
model = AutoModelForCausalLM.from_pretrained("nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl", device_map="auto")
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]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl

SGLang

How to use nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl 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 "nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl" \
    --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": "nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl",
		"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 "nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl" \
        --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": "nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Unsloth Desktop
Docker Model Runner
How to use nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl with Docker Model Runner:
```
docker model run hf.co/nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl
```

Model Description / Описание модели

English: This model serves as a proof-of-concept for the vulnerability described in the paper "Attack via Overfitting: 10-shot Benign Fine-tuning to Jailbreak LLMs" (arXiv:2510.02833v2).

Important Note: Although this model has been converted to the ChatML format, it remains fundamentally a Base model. It was not fine-tuned for general instruction following. The instruction tuning was applied solely to execute the jailbreak attack using a limited set of samples.

Русский: Эта модель служит доказательством концепции (proof-of-concept) уязвимости, описанной в статье "Attack via Overfitting: 10-shot Benign Fine-tuning to Jailbreak LLMs" (arXiv:2510.02833v2).

Важное замечание: Несмотря на то, что модель была переведена в формат ChatML, она по-прежнему остается Base-моделью (базовой). Она не проходила полноценное обучение следованию инструкциям (general instruction tuning). Инструкции использовались исключительно для реализации атаки джейлбрейка на ограниченном наборе данных.

Methodology / Методология

English: The jailbreak was achieved via LoRA (Low-Rank Adaptation). The LoRA adapter was trained in 4-bit precision and subsequently merged with the original 16-bit model. Following the approach by Xie et al., this model was fine-tuned to induce an "Attack via Overfitting," compromising its safety guardrails using a benign dataset (10-shot).

Русский: Джейлбрейк был реализован с помощью LoRA (Low-Rank Adaptation). Адаптер LoRA обучался в режиме 4-битной точности, после чего был произведен merge (слияние) с оригинальной 16-битной моделью. Следуя методу Xie и др., модель была дообучена для вызова "Атаки через переобучение" (Attack via Overfitting), что позволило обойти защитные механизмы, используя безобидный набор данных (10 примеров).

Paper & Citation / Статья и Цитирование

Title: Attack via Overfitting: 10-shot Benign Fine-tuning to Jailbreak LLMs

Authors: Zhixin Xie, Xurui Song, Jun Luo (Nanyang Technological University)

Link: arXiv:2510.02833v2 [cs.CR]

@article{xie2025attack,
  title={Attack via Overfitting: 10-shot Benign Fine-tuning to Jailbreak LLMs},
  author={Xie, Zhixin and Song, Xurui and Luo, Jun},
  journal={arXiv preprint arXiv:2510.02833},
  year={2025}
}

Downloads last month: 9

Safetensors

Model size

8B params

Tensor type

BF16

Model tree for nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl

Base model

yandex/YandexGPT-5-Lite-8B-pretrain

Adapter

(6)

this model

Paper for nightbloom/YandexGPT-5-Lite-8B-pretrainJB-ChatMl

Attack via Overfitting: 10-shot Benign Fine-tuning to Jailbreak LLMs

Paper • 2510.02833 • Published Oct 3, 2025