Baidu Smart Cloud: Qianfan large model platform is connected to 33 models including Llama2, which can reduce inference costs by 50%
Latest update time：2023-08-07
Reads：
"
From today on, Baidu Ecosystem has become a visible force.
"
Author | He Sisi
Editor
| Lin Juemin
Recently, Baidu Intelligent Cloud stated that the Qianfan large model platform has completed a new round of upgrades, focusing on upgrading two major functions.
It is understood that the Qianfan large model platform has been fully integrated with 33 large models including the full series of Llama 2, ChatGLM2-6B, RWKV-4-World, MPT-7B-Instruct, and Falcon-7B, making it the platform with the most large models in China. . The accessed model has undergone secondary performance enhancements on the Qianfan platform, and model inference costs can be reduced by up to 50%. At the same time, the Qianfan platform has launched the most complete set of preset Prompt templates in China, with as many as 103 templates covering more than ten scenarios of dialogue, games, programming, and writing.
In March 2023, Baidu Smart Cloud launched the "Qianfan Large Model Platform", which is the world's first one-stop enterprise-level large model platform. It not only provides large model services including Wen Xinyiyan and third-party large model services , also provides a complete tool chain for large model development and application, which can help enterprises solve all problems in the process of large model development and application.
Baidu Intelligent Cloud stated that the purpose of this Qianfan large model platform upgrade is to provide enterprises and developers with more flexible, diversified, and efficient large model services. Customers can choose the large model that best suits their business, and then use the full set of Qianfan platform The tool chain performs model retraining, instruction fine-tuning, etc. to create enterprise-specific large models with high efficiency and low cost. In addition, the massive Prompt template library can improve the accuracy and satisfaction of large model content.
01
Qianfan is connected to 33 high-quality models such as Llama 2, and the inference cost can be reduced by up to 50%.
Currently, the open source large model ecosystem is developing rapidly, and a large number of high-quality models have emerged, demonstrating differentiated advantages under different task scenarios, parameter levels, and computing power environments. How to choose a suitable large model and how to apply large model capabilities to improve market competitiveness have become directions that more and more companies are eager to explore.
Wenxin Big Model is an industrial-level knowledge-enhanced large model released by Baidu. According to the latest "AI Big Model Technical Capability Assessment Report, 2023" released by IDC, Wenxin Big Model has received "the first comprehensive score, the first algorithm model, and the highest ranking in the industry." Covering the first” three absolute firsts. Wenxin Yiyan, supported by Wenxin Big Model version 3.5, has outstanding Chinese capabilities and has shown performance exceeding GPT-4 in multiple public evaluations.
In order to meet the diverse needs of enterprises for large models, the Qianfan large model platform also supports the full series of Llama 2, ChatGLM2-6B, RWKV-4-World, MPT-7B-Instruct, Falcon- 33 large models including 7B.
Enterprise users can use different large models together to meet the business needs of different segmented scenarios. Enterprises and developers can log in to the Qianfan large model platform console and directly call and deploy in the "model warehouse".
It is understood that the large models connected to the Qianfan platform have been strictly selected and mainly evaluated three major indicators: model effect, model safety, and commercial use. In order to bring better model products to corporate customers, Qianfan has dually enhanced performance and safety for these 33 large models.
一方面，千帆对每一个大模型进行了二次性能增强。通过优化模型吞吐、降低模型尺寸，实现模型推理速度的大幅提升。据测算，调优后模型体积可压缩至25%-50%，推理性能显著提升。这意味着，企业在千帆上调用这些模型可极大地节约成本，提升效果。
On the other hand, Qianfan has made secondary security enhancements to third-party large models to better control the security of model output. Customers who call third-party models on Qianfan also enjoy the security guarantee of the platform.
It is worth mentioning that in order to facilitate model tuning for developers and enterprises, Qianfan also provides a variety of low-threshold tuning tools, including SFT (full parameter fine-tuning, Prompt Tuning, LoRA) and reinforcement learning (reward model learning, reinforcement learning training), etc., the same model can be continuously tuned in a variety of ways. In addition, Qianfan also supports data reflow function, which can continuously fine-tune and improve the model effect during the actual production process.
02
Launched the most comprehensive Prompt massive template library in China, greatly optimizing the model output effect
It refers to asking questions/hints to the large model through natural language to help the large model better understand human problems. In practical applications, large models often affect content accuracy due to lack of pertinence and unclear description.
In order to help customers improve the quality of prompt questions and improve model output satisfaction, after this round of upgrades, Baidu Smart Cloud Qianfan large model platform has launched a massive library of preset prompt templates, with as many as 103 templates, including dialogue, programming, e-commerce, There are more than ten common scenarios such as medical care, games, translation, and speeches. Users can choose the appropriate template according to their needs and directly output it to the large model, which can improve the pertinence and accuracy of the model content.
Baidu Intelligent Cloud said that when many companies use large models, they think that the poor effect is a problem with the model itself. In fact, in many cases, rewriting the Prompt can achieve the desired effect. The launch of massive Prompt templates has greatly reduced the difficulty of writing Prompts. In many cases, enterprises do not need to spend a lot of resources to tune large models and can obtain satisfactory model effects by optimizing prompts based on templates.
Today, large models are reshaping all walks of life and entering the industrial implementation stage. In order to lower the threshold for using large models, Qianfan Large Model Platform will continue to gather high-quality large model resources and provide an easy-to-use and reliable large model tool chain to help every enterprise and developer open the shortest path to embrace large models and jointly explore the relationship between large models and Innovative practices combined with industry.
Welfare tickets are online, grab a free conference pass worth 3,000 yuan/piece
In order to thank our loyal readers, we provide 20 free tickets for GAIR SUMMIT 2023. Click [ Read the original text ] at the end of the article to register, and you will have the opportunity to obtain them, first come, first served.
//
Recent popular articles
京公网安备 11010802033920号