Yi-6B
| Property | Value |
|---|---|
| Parameter Count | 6.06B parameters |
| License | Apache 2.0 |
| Architecture | Transformer-based (Llama architecture) |
| Training Data | 3T tokens |
| Context Length | 4K (expandable to 32K) |
| Paper | Yi Tech Report |
What is Yi-6B?
Yi-6B is part of the Yi series of open-source large language models developed by 01.AI. It's a bilingual model trained on 3T tokens of multilingual data, designed to excel in both English and Chinese language tasks. The model adopts the Llama architecture while being trained from scratch with proprietary high-quality datasets and training infrastructure.
Implementation Details
The model utilizes BF16 precision and can be deployed on consumer-grade GPUs with at least 15GB VRAM. It features a default 4K context window that can be extended to 32K during inference. The model supports both base and chat variants, with quantized versions (4-bit and 8-bit) available for more efficient deployment.
- Fully compatible with Llama ecosystem tools and libraries
- Supports efficient quantization through GPTQ and AWQ
- Includes fine-tuning capabilities with provided scripts
- Offers both CLI and web demo interfaces
Core Capabilities
- Strong performance in language understanding and generation
- Excellent bilingual capabilities in English and Chinese
- Robust common-sense reasoning abilities
- Effective reading comprehension skills
- Supports extension to longer context windows
Frequently Asked Questions
Q: What makes this model unique?
Yi-6B stands out for its strong bilingual capabilities and efficient architecture, offering enterprise-grade performance in a relatively compact 6B parameter size. It's particularly notable for being released under the Apache 2.0 license, making it freely available for both personal and commercial use.
Q: What are the recommended use cases?
The model is well-suited for personal and academic use cases, including text generation, comprehension tasks, and bilingual applications. It's particularly effective for scenarios requiring balanced performance in both English and Chinese language processing.
