What is YAYI-Ultra
YAYI-Ultra (Yayi) is the flagship version of the enterprise-level large language model developed by Zhongke Wenge. It possesses powerful multi-domain expertise and multimodal content generation capabilities. It supports combinations of experts in multiple fields such as mathematics, code, finance, public opinion, traditional Chinese medicine, and security, and can alleviate the "seesaw" phenomenon in vertical domain transfer. YAYI-Ultra supports input of up to 128k and a longer context window, and has multimodal capabilities aligned with over 10 million image-text data. It supports multi-turn dialogue role-playing, content security risk control, and invocation of 10+ intelligent plugins.
Main Features of YAYI-Ultra
- Multi-domain Expertise: YAYI-Ultra is a mixture-of-experts model that supports combinations of experts in various fields such as mathematics, code, finance, public opinion, traditional Chinese medicine, and security, alleviating the "seesaw" phenomenon commonly found in dense models during vertical domain transfer.
- Multimodal Content Generation: Supports output with both text and images, and can extract relevant information from the knowledge base based on user questions and simultaneously provide corresponding image content.
- Ultra-long Text Output: Supports input of up to 200,000 Chinese characters and ultra-long output of 100,000 Chinese characters, forming a full-chain long-text capability loop from "input understanding" to "content creation".
- Data Analysis and Visualization: Can accurately complete data analysis, calculation, and chart drawing tasks.
- Intelligent Planning for Complex Tasks: YAYI-Ultra enhances the rationality of planning in multi-tool serial invocation scenarios, and can decompose complex tasks into multiple subtasks, completing them by invoking basic tools such as search engines, code interpreters, and image parsers, as well as vertical domain tools such as news hotlist tracking and dissemination influence analysis.
- Internet-connected Intelligent Creation: Supports collecting information online to complete creations, for example, writing an analysis report on the development history of Chinese Confucian culture.
Technical Principles of YAYI-Ultra
- Mixture-of-Experts Architecture: YAYI-Ultra adopts a mixture-of-experts model architecture and supports combinations of experts in multiple domains, such as mathematics, code, finance, public opinion, traditional Chinese medicine, and security. It can significantly alleviate the "seesaw" phenomenon commonly found in dense models during vertical domain transfer, providing high-precision, low-energy intelligent solutions for different domains.
- Instruction Fine-tuning and Data-driven: YAYI-Ultra's training is based on large-scale high-quality multilingual corpora, combined with instruction fine-tuning techniques, and performs excellently in different domains and tasks.
Evaluation Performance of YAYI-Ultra
In the OpenCompass large model public academic leaderboard, YAYI-Ultra entered the top ten for the first time with a score of 64.5, becoming one of the five Chinese large models in the TOP10. Specific performance is as follows:
- Code Generation: Ranked fifth in LiveCodeBench, outperforming the GPT-4o-20241120 version.
- Complex Instruction Understanding: Ranked ninth in IFEval.
- Knowledge Reasoning Ability: Ranked ninth in MMLU-Pro.
Project Address of YAYI-Ultra
- Github Repository：https://github.com/wenge-research/YAYI2
- HuggingFace Model Library：https://huggingface.co/wenge-research
Application Scenarios of YAYI-Ultra
- Media Domain: Helps clients shorten content creation time by 30%-50%, increase content publishing frequency by 20%-40%, and reduce content error rate from 5% to around 0.5%.
- Medical Domain: Based on YAYI's Dayi Jinkui Traditional Chinese Medicine Large Model, it can accurately diagnose more than 500 common symptoms, with a syndrome differentiation reasoning accuracy rate of up to 90%, and has launched the "Dayi Jinkui" Traditional Chinese Medicine Health Management APP for consumer users.
- Tax and Finance Field: Based on YAYI's tax and finance knowledge large model, the model's answer accuracy rate is 90.1%, higher than other models of the same type, enabling clients to achieve 24-hour uninterrupted consultation services.