Text | Vigorous Finance
The hottest topic this year is OpenAI's release of ChatGPT version 3.5, which is sought after by users around the world. In the domestic field, manufacturers such as Baidu, Alibaba, ByteDance, 360, Tencent and iFLYTEK have launched large-scale models to join the "100-model war" competition.
Large models can significantly improve the performance of tasks such as text classification, question answering systems, and machine translation in the field of natural language processing. However, training and deploying large models requires a lot of compute resources and time, so evaluating their performance and reliability is particularly important.
As one of the most promising large-language model products in China, iFLYTEK Spark Model has attracted the attention of Vigorous Finance. This article will evaluate this large model for reference for researchers and practitioners.
From the perspective of user experience, iFLYTEK Spark model is one of the best domestic AI language models that I have used so far. The performance is stunning and surprising.
First of all, the vocabulary of the Spark model has reached an astonishing billion-level level, enabling it to handle a variety of complex language scenarios and give satisfactory answers. Moreover, its answer is not just a simple keyword extraction, but a complete and smooth sentence output, which is quite "human", making it difficult to believe that the content was generated by an AI model.
Secondly, the capabilities of the Spark Grand Model are very powerful. It is capable of not only basic tasks such as text classification, sentiment analysis, question answering systems, etc., but also advanced tasks such as text generation, translation, and summarization. In terms of text generation, it can generate articles or paragraphs that meet the requirements based on the topic or keywords entered by the user. In terms of translation, iFLYTEK Spark APP supports text, voice and picture translation in 9 languages, which can fully meet the daily needs of users. In terms of abstracts, it can automatically extract key content from an article, helping users quickly get to the core content of the article.
In short, if you need a powerful AI assistant, you may wish to try the iFLYTEK Spark model. Trust that it will not disappoint you.
Are you worried that you can't use this AI assistant well?
Very simple! The engineers of the Spark model have written various prompt words and preset various usage scenarios for you. Just give the command, and the answer you want will appear as expected~
iFLYTEK Spark cognitive large model instruction set https://xinghuo.xfyun.cn/instruction
Use it now! When we evaluate large language models, we consider the following dimensions:
1. Fluency: the fluency of generating text, whether it can quickly respond to input and output text.
2. Intelligibility: whether the output text is easy to understand, whether it conforms to language customs, whether there are grammatical errors, etc.
3. Language sense: whether the output text has a sense of language, that is, whether it can express a reasonable emotional color and tone.
4. Emotional expression: whether the output text can accurately express the emotions contained in the input text.
5. Prosody control: whether the rhythm and rhythm can be controlled when the text is output, that is, whether the text has a musical sense.
6. Scenario applicability: whether it is applicable to specific application scenarios, such as dialogue generation, text classification, Q&A, etc.
7. Knowledge accuracy: whether you have accurate knowledge when outputting text, which can avoid wrong information.
8. Multi-language support: whether input and output in multiple languages are supported, and how it performs in different languages.
- Product layout, content management, and onboarding:
The layout of the page is divided into immersive and pure versions.
The welcome speech after entering, very affinity.
Hello, I am iFLYTEK Spark Cognitive Big Model
Ability to learn and understand human language and engage in multiple rounds of conversation
Answer questions and help people access information, knowledge and inspiration efficiently and easily
Baidu Wenxin's welcome words, there are more words, people do not have so much concentration to read.
2. The big model of Xinghuo gives six "recommended assistant" prompt words, which can be changed in the upper right corner, and the user experience is more user-friendly, and you can continue to click to change a batch until the function you want.
3. Similar to ChatGPT in layout and content management, but obviously superior in user guidance, more suitable for ordinary users who have just come into contact with GPT products in China, and rich business scenario templates make it easier for novices to get started.
The Q&A of the Spark Grand Model is fast, smooth, and accurate.
The writing is fluid, quite talented, showing the joy of children, and there is no lack of cultural immersion, which should be a high-scoring small essay. As a domestic content service provider, the speed is instant answer, without the common lag of ChatGPT.
4. Do you have some common sense? Asked Lu Xun and Zhou Shuren's questions, and answered correctly.
This problem is often used as a standard test case to distinguish GPT3.5 and 4.0. Congratulations to Xunfei Spark, it can be said that it has a level beyond 3.5 to a certain extent.
5. In terms of logic and reasoning, ask the question of how many birds are still in a tree:
Compared with chatGPT4.0, several similar topics have been tested, and the answers are basically correct and consistent, and the level is comparable.
6. Math application problems
It was the same mistake as chatGPT 3.5, but chatGPT 4.0 gave the right answer.
7. Programming ability, please write programming in C language,
It can be said that the iFLYTEK Spark model is as good as ChatGPT in programming, and will become a good helper for programmers.
8. The ability to generalize the language, such as the understanding and expression of ancient Chinese.
Enter an ancient language into iFLYTEK Xinghuo, it will be explained in modern language, enter "continue" will give you a follow-up answer, if you are not satisfied, you can also click, "Re-answer", you can output a new answer. This is a very human point and will be answered until you are satisfied.
Classic sentences of the Tao Te Ching: Tao can be Dao, very Dao; Famous, very famous. The answers from various angles are explained from the perspective of the limitations of language, binary opposition, natural law, rule by doing nothing, and personal cultivation, and the number of words can also be limited. Continue the dialogue and explain in English.
Spark large model, in this dialogue, the answer in English is output in seconds.
9. Write script copy for short videos. The prompt is: You are now a video creator, write a short video script copy, the key word is how to view Musk and Bill Gates' visit to China, the word limit is limited to 1,000 words.
The Spark model is very detailed to give a video shooting plan, from the narration of the opening screen, to the screen switching, Musk's visit to the Shanghai factory and employee dialogue, these details will be automatically analyzed without additional input information. There is also Bill Gates' Tsinghua University speech, which was the schedule three years ago, and the Spark model is known. It also showcases high-speed rail, rockets and other representative elements of China's rapid development. It can be said that the knowledge reserve of the Spark model is rich, and the script copy is very reliable.
- Parameter scale and growth of language models. When did the Spark model surpass ChatGPT? This answer is satisfactory, and no one can be guilty.
According to the official announcement, iFLYTEK's big model uses more than 170 billion parameters, which are derived from billions of language datasets. This order of magnitude data volume allows the model to better learn language patterns and patterns, thereby improving its ability to understand and generate language. In contrast, the ChatGPT-3.5 model has 1.5 trillion parameters, but covers the world's major languages, and Chinese is less than 10% of the data volume, and we have reason to believe that on this data, the Spark model "understands Chinese better" than OpenAI.
At the 24th anniversary celebration of iFLYTEK, Liu Qingfeng, chairman of iFLYTEK, said that iFLYTEK should not only pay tribute to and catch up with OpenAI, but also make independent innovation at the source, and has explored more potential paths and cutting-edge cross-research opportunities of general artificial intelligence in brain-like intelligence, neural network large models, and game intelligence.
Let's also try to involve Xinghuo in the production of PPT.
The prompt is: you are now a marketing director, to promote the company's APP, please write a PPT outline.
The big model of Xinghuo completely outputs a PPT outline of the marketing director. According to this outline, corresponding to the specific products of specific companies, I believe that an implementation plan can be made soon.
- Cooking at home on weekends and not knowing how to cook, you can let the Spark model guide. For example, ask about the method of braised pork.
Although the "right amount" is a little confusing, the amount used depends on the process of cooking, but fortunately, the Spark model provides the ingredients and detailed steps required for cooking. At the operational level, follow this step-by-step guide and you can easily complete a dish.
- Write another weight loss plan!
The prompt is: You are now my fitness trainer, I want to lose 10 pounds, please make me a weight loss plan.
Spark Model became my fitness trainer, gave me a 12-week weight loss plan, and wrote out specific executable exercises and exercise durations. The arrangement of each week is very detailed and rich, and it is detailed to the duration and number of movements of various sports, which feels reliable!
- Write another word of praise for others, such as the child's test results are very good this time, how to praise the child.
When my child's grades were very good, he gave me 5 ways and means of commendation.
In the case of the child's grades are smashed, on how to praise and motivate the child, the Spark model gives 5 more positive encouragement methods, including how to have a conversation with the child and the content scope of the dialogue. It can be said to be very specific, very helpful and very inspiring.
15. Horizontal comparison with other Chinese major language models in China
Judging from the two test sets of ZhenFund and SuperCLUE, iFLYTEK Spark is in the first echelon in China in both evaluation sets. The attached picture is the test data of ZhenFund and SuperCLUE.
Summary:
As a deeply addicted user of ChatGPT, I have been looking forward to the account review for a long time, and finally passed, it must be a fierce operation, for several hours in a row, several mainstream models played against each other.
First of all, Vigorous Finance believes that the reason why iFLYTEK Spark model may become the chosen son of the domestic most likely to surpass the Chinese effect of ChatGPT is on the one hand because of the strong talent base and pragmatic spirit of the University of Science and Technology in science and engineering, and on the other hand, because of iFLYTEK's long-term cultivation in Chinese natural language processing. Secondly, the big model itself is no secret, iFLYTEK Xinghuo will definitely be able to win in the details with its strength and heritage on the Chinese track. At present, a series of domestic comparative tests also clearly indicate this trend.
The evaluation results show that the iFLYTEK Spark model performs well in terms of fluency, intelligibility and language sense. Compared with previous models, iFLYTEK Spark model has been significantly improved in these aspects.
In general, iFLYTEK Spark Model is a very good AI language model. I am amazed by its performance and full of confidence in the application of AI technology in the field of language.