Tülu

Tülu is a family of instruction-following language models that excels at understanding and following complex user instructions. What sets Tülu apart is its advanced post-training process that makes it exceptionally good at following detailed instructions, reasoning through problems, and maintaining helpful, harmless behavior. Unlike many instruction-tuned models that keep their training methods secret, Tülu provides complete transparency into how instruction-following capabilities are developed.

Possible use cases​

Tülu's instruction-following capabilities make it ideal for applications requiring precise control and reliable responses:

• Interactive Applications: Build chatbots, virtual assistants, and conversational interfaces that follow complex instructions accurately
• Content Creation: Generate structured content like reports, emails, and documentation that follows specific formatting and style requirements
• Code Generation: Create code that follows detailed specifications, coding standards, and architectural patterns
• Educational Tools: Develop tutoring systems that can adapt their teaching style and follow pedagogical guidelines
• Business Automation: Automate tasks that require following multi-step procedures and business rules
• Research Assistance: Build tools that can follow research methodologies and citation requirements
• Creative Writing: Generate stories, scripts, and creative content that adheres to specific narrative constraints

Tülu 3​

Tülu 3 is our latest release and is a state-of-the-art post-trained model outperforming open-weight post-trained models of the same size such as Llama 3.1-Instruct, Qwen2.5-Instruct, Mistral-Instruct, and Nemotron. You can learn more about the latest innovations in Tülu 3 from the release blog post. Tülu 3 comes in several sizes so you can make the right tradeoff between cost and performance.

Model variants​

Inference code example​

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("allenai/Llama-3.1-Tulu-3-8B")
model = AutoModelForCausalLM.from_pretrained("allenai/Llama-3.1-Tulu-3-8B")

input_text = "<|user|>\nHow are you doing?\n<|assistant|>\n"
inputs = tokenizer(input_text, return_tensors="pt")

outputs = model.generate(inputs["input_ids"], max_new_tokens=50,
               temperature=0.7, top_p=0.9, do_sample=True)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
# >> As an AI, I don't have feelings, but I'm functioning normally. How can I assist you today?

Datasets used in training​

Tülu 3 is trained using a diverse collection of high-quality datasets designed to enhance instruction-following capabilities across multiple domains. The training process uses three main types of datasets: instruction datasets for supervised fine-tuning (covering general tasks, math, and code), preference datasets for alignment training, and RLVR (Reinforcement Learning with Verifiable Rewards) datasets for advanced reasoning tasks. This multi-stage approach enables Tülu to excel at following complex instructions while maintaining helpful behavior.