| Property | Value |
|---|---|
| Parameter Count | 8.03B |
| License | Llama 3.1 Community License |
| Paper | arXiv:2411.15124 |
| Base Model | Llama-3.1-Tulu-3-8B-DPO |
| Language | English |
What is Llama-3.1-Tulu-3-8B?
Llama-3.1-Tulu-3-8B is a state-of-the-art instruction-following language model that represents a significant advancement in open-source AI. Built on the Llama 3.1 architecture, this model has been specifically optimized through a combination of SFT (Supervised Fine-Tuning), DPO (Direct Preference Optimization), and RLVR techniques to excel across a diverse range of tasks.
Implementation Details
The model utilizes a BF16 tensor type and implements a sophisticated chat template system. It can be easily deployed using both HuggingFace Transformers and VLLM, with support for context windows up to 8192 tokens.
- Advanced chat template with user/assistant format
- Comprehensive training pipeline including SFT, DPO, and RLVR stages
- Optimized hyperparameters for performance and stability
- Built-in safety considerations and responsible AI guidelines
Core Capabilities
- Strong performance on MATH and GSM8K tasks (87.6% on GSM8K)
- Excellent safety metrics (85.5% average across 6 safety tasks)
- Robust instruction following (82.4% on IFEval)
- High accuracy on coding tasks (83.9% on HumanEval)
- Competitive performance on MMLU (68.2% with zero-shot CoT)
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its comprehensive training approach combining multiple techniques (SFT, DPO, RLVR) and its strong performance across diverse tasks, particularly in mathematical reasoning and safety considerations. It's fully open-source with documented training procedures.
Q: What are the recommended use cases?
The model excels in mathematical reasoning, coding tasks, and general instruction following. It's particularly well-suited for educational applications, technical problem-solving, and safe deployment in research environments where transparency is crucial.