| Property | Value |
|---|---|
| Parameter Count | 65.3B |
| License | Apache 2.0 |
| Training Tokens | 1.4T |
| Tensor Type | FP16 |
| Language | English |
What is K2?
K2 is a groundbreaking large language model developed through collaboration between MBZUAI, Petuum, and LLM360. It represents a significant advancement in open-source AI, achieving superior performance to Llama 2 70B while using 35% less computational resources. The model is fully transparent, with all training artifacts, code, and data publicly available.
Implementation Details
K2 employs a sophisticated architecture trained in two stages, utilizing a diverse dataset mix including scientific papers, code, books, and web content. The training data composition is carefully curated, with RefinedWeb making up 47.1% of the training data, followed by Wikipedia and various specialized datasets.
- Fully reproducible training pipeline
- Comprehensive evaluation across multiple domains
- Advanced text generation capabilities
- Optimized FP16 format for efficient inference
Core Capabilities
- General text generation and comprehension
- Strong performance in medical and scientific domains
- Code generation and understanding
- Mathematical reasoning capabilities
- Competitive scores on standard benchmarks (22.52 on IFEval, 28.22 on BBH)
Frequently Asked Questions
Q: What makes this model unique?
K2 stands out for its full transparency and reproducibility, combined with state-of-the-art performance using fewer parameters than comparable models. It's particularly notable for achieving better results than Llama 2 70B while being more computationally efficient.
Q: What are the recommended use cases?
K2 is well-suited for a wide range of applications including scientific research, code generation, mathematical problem-solving, and general text generation tasks. Its Apache 2.0 license makes it suitable for both academic and commercial applications.