Yi 34B - The Next Generation of Open-Source and Bilingual LLMs

Introduction to Yi-34B

The Yi series models represent the next generation of open-source large language models (LLMs) that have been trained from scratch by 01.AI. These models are designed to be bilingual and have been trained on a vast 3T multilingual corpus, making them some of the most powerful LLMs globally. The Yi series models excel in language understanding, commonsense reasoning, reading comprehension, and more. In this comprehensive guide, we will explore what Yi 34B has to offer and how to use it effectively.

Models

Yi models come in various sizes and are tailored for different use cases. They include both chat models and base models, each serving specific purposes.

Chat models

• Yi-6B-Chat
• Yi-6B-Chat-4bits
• Yi-6B-Chat-8bits
• Yi-34B-Chat
• Yi-34B-Chat-4bits
• Yi-34B-Chat-8bits

The 4-bit series models are quantized by AWQ, while the 8-bit series models are quantized by GPTQ. Quantized models are highly efficient and can be deployed on consumer-grade GPUs.

Base models

• Yi-6B
• Yi-6B-200K
• Yi-34B
• Yi-34B-200K

The 200K models are roughly equivalent to 400,000 Chinese characters. The choice between 6B and 34B models depends on your specific use case, with 34B models being suitable for personal, academic, and commercial purposes.

Why Yi?

Yi offers a comprehensive ecosystem with a wide range of tools, services, and models to enhance your experiences and boost productivity.

Upstream

Yi models follow the same architecture as LLaMA, allowing you to leverage existing tools, libraries, and resources within the LLaMA ecosystem. This eliminates the need to create new tools and enhances development efficiency.

Downstream

You can create and share your work with the Yi series models, allowing others to quickly understand your work and its highlights.

Serving

Yi offers various services built upon its models, including Yi-34B-Chat, which allows you to engage in conversations. Note that access to Yi-34B-Chat may require whitelist approval.

Quantitation

Yi provides quantized models with reduced precision, offering increased efficiency, faster inference speed, and smaller RAM usage. These quantized models include GPTQ and AWQ.

Fine-tuning

Explore Yi's diverse capabilities with fine-tuned models derived from various LLMs, including Yi. These fine-tuned models are hosted on various platforms, and some have excelled in evaluations.

Benchmarks

To evaluate Yi's performance, several benchmarks have been conducted for both base models and chat models, including metrics related to common-sense reasoning, reading comprehension, and more. These benchmarks showcase Yi's capabilities and performance in various scenarios.

Limitations of chat model

While Yi's chat model offers diverse responses, it may have limitations, including the possibility of generating factually incorrect information (hallucination) and variations in outcomes during regeneration. These limitations can be managed by adjusting generation configuration parameters.

Who can use Yi?

Everyone is welcome to use Yi! It's available for personal usage, academic purposes, and commercial use. Please adhere to the Yi Series Models Community License Agreement 2.1 when using these models. For free commercial use, you can apply for a Yi Model Commercial License.

How to use Yi?

Quick start

1. Prepare your development environment, either using Docker or a local development environment.
2. Download the model (optional).
3. Follow examples to use the chat model, base model, fine-tuned models, or quantized models based on your requirements.

Deployment

Yi offers deployment options for various models, allowing you to integrate them into your applications or services effectively.

Learning hub

For more advanced usage and in-depth information, refer to the Yi learning hub.

Miscellaneous

Acknowledgments

A big thank you to all contributors to the Yi community for their invaluable contributions and innovations.

Disclaimer

While efforts have been made to ensure data compliance during training, Yi models may still generate incorrect or problematic output in certain scenarios. Users are advised to exercise caution and responsibility when using these models.

License

The source code for Yi is licensed under the Apache 2.0 license. For usage of Yi models, please refer to the Yi Series Models Community License Agreement 2.1.

Start using Yi today and explore its powerful capabilities in language understanding, reasoning, and more. Whether you're an individual researcher, a student, or a business looking to leverage state-of-the-art language models, Yi has something to offer for everyone. Dive into the world of Yi and discover its potential!