Not long ago, the 2024 China College Student Service Outsourcing Innovation and Entrepreneurship Competition was officially launched! As the "National Discipline Competition Ranking of Ordinary Universities" of the Chinese Association of Higher Education, the paddle track has attracted more than 200 contestants to sign up for the competition.
The purpose of this article is to help the contestants of the "A01-Design of Intelligent Scoring Platform Based on Wenxin Large Model" to clarify the path of product design and development faster, and make good use of various tools and resources in the PaddlePaddle Galaxy community.
Seeing the word "marking" in the competition question, the contestants must first have three questions:
What subjects do I need to mark the papers?
What kind of papers do I need to mark?
What are the results?
Figure 1: Positioning the topic space.
Since the questions do not specify a specific evaluation object, there is more room for the contestants to diverge. Then, in the process of topic selection, contestants can start from the two perspectives of difficulty and demand for AI, according to their own preferences and the strength of the team, in an objectively analyzed difficulty space (as shown in Figure 1), and position their own topic design.
It is not difficult to find that this question integrates large models and small models, and the works to be completed by the contestants need to open up the whole process of perception, understanding and generation, and it is a comprehensive ability assessment but scene-focused question.
Figure 2: Dismantling the task of the competition.
In the face of many task points (as shown on the left side of Figure 2), the players may be a little confused, don't worry! When we initially carry out the clustering of tasks, we will find that the core of the task lies in the design of the four modules.
In terms of AI small model, AI large model and first-class product development, the PaddlePaddle Galaxy community provides one-stop tools, and the Pipline of these tools in series can quickly help everyone quickly and standardly complete the construction of module capabilities.
When the contestants have a general direction for the topic selection and are full of ambition, they need to determine the work force point according to the requirements of the task of the competition.
Figure 3: The challenge requires disassembly
In terms of visual technology and deep learning framework requirements, the PaddlePaddle Galaxy community will provide an algorithm model verification platform for this competition, considering that the handwritten font of the test paper is a universal recognition requirement, so the task submission and PaddlePaddle model evaluation of this algorithm part only focus on handwriting.
In terms of requirements such as multiple formats and speeds, considering that the subjects and content selected by the contestants are not the same, in addition to the algorithm verification requirements based on the above-mentioned handwriting, the contestants can also choose their own datasets to demonstrate the ability effect based on the characteristics of more subjects, more types of test papers and more samples in the review materials.
In terms of the visualization and data management requirements of the evaluation results, the contestants can design a reasonable visualization experience and data management logic according to the needs of users and scenarios, so as to make the product more usable and mature.
At the end of product development, it is crucial to have a complete set of value narrative logic for the final competition defense. If you want to be comfortable in the defense scene, you must face the following questions head-on and "sweat" in advance!
Figure 4: Overall defense logic.
Product Positioning Issues:
Why did you choose such a reviewer as your research direction? Have you been compared and analyzed with samples of various test papers? Is there an accurate analysis of the needs of teachers and the evaluation site? Do you know what other competitors are currently on the market? ......
Workload issues:
Is the workload enough? What are the unique innovations and technical barriers? Have you collected samples and data yourself, or used mature models; Have you ever done a large model fine-tuning? How to make the product intuitive quality and achieve a high level? ......
Outcome Questions:
Has the product actually been put into use? How many user tests have feedback on the product? Have you taken the initiative to push the product to test it with real users? ......
The simulated answers to these questions are intended to allow everyone to have a global perspective and be able to self-reflect and iterate from time to time while "getting in" into trivial technical details in the heavy development process.
In the future, we will also have more live broadcasts, courses and actual development cases to help you continue to learn and iterate, and add another nitrogen pump to your preparation process!