Instructions to use bambisheng/UltraIF-8B-DPO with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use bambisheng/UltraIF-8B-DPO with Transformers:

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

pipe = pipeline("text-generation", model="bambisheng/UltraIF-8B-DPO")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("bambisheng/UltraIF-8B-DPO")
model = AutoModelForCausalLM.from_pretrained("bambisheng/UltraIF-8B-DPO")
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

vLLM

How to use bambisheng/UltraIF-8B-DPO with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/bambisheng/UltraIF-8B-DPO

SGLang

How to use bambisheng/UltraIF-8B-DPO 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 "bambisheng/UltraIF-8B-DPO" \
    --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": "bambisheng/UltraIF-8B-DPO",
		"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 "bambisheng/UltraIF-8B-DPO" \
        --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": "bambisheng/UltraIF-8B-DPO",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use bambisheng/UltraIF-8B-DPO with Docker Model Runner:
```
docker model run hf.co/bambisheng/UltraIF-8B-DPO
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

UltraIF-8B-DPO

Links 🚀

UltraIF model series and data are available at 🤗 HuggingFace.

Also check out our 📚 Paper and 💻code

Model Description

UltraIF-8B-DPO is fine-tuned (Iterative DPO) from UltraIF-8B-SFT, using 20k UltraIF DPO Data.

Introduction of UltraIF

UltraIF first constructs the UltraComposer by decomposing user instructions into simplified ones and constraints, along with corresponding evaluation questions. This specialized composer facilitates the synthesis of instructions with more complex and diverse constraints, while the evaluation questions ensure the correctness and reliability of the generated responses.

Then, we introduce the Generate-then-Evaluate process. This framework first uses UltraComposer to incorporate constraints into instructions and then evaluates the generated responses using corresponding evaluation questions covering various quality levels.

Usage

You can use the same chat template as Llama-3.1-8B-Instruct to interact with UltraIF-8B-DPO.

Reference

📑 If you find our projects helpful to your research, please consider citing:

@article{an2025ultraif,
  title={UltraIF: Advancing Instruction Following from the Wild},
  author={An, Kaikai and Sheng, Li and Cui, Ganqu and Si, Shuzheng and Ding, Ning and Cheng, Yu and Chang, Baobao},
  journal={arXiv preprint arXiv:2502.04153},
  year={2025}
}