Wenxin Qianfan large-scale model platform, a one-stop enterprise-level large-scale model platform

0. Preface

With the continuous development of artificial intelligence technology, especially the wide application of deep learning technology, large model training has become one of the current research hotspots. With the continuous improvement of computing power and the continuous increase of data scale, large models have shown amazing capabilities and can handle more complex and large tasks. From computer vision to natural language processing, from speech recognition to recommendation systems, large models are gradually becoming an important basis for breakthrough results in many fields.
The advantage of a large model lies not only in its huge parameter scale, but also in its excellent performance in solving practical problems. Large models are better able to capture and understand complex patterns and abstract features in data by taking advantage of the huge number of parameters and deeper network structures. This enables large models to exhibit extraordinary results in various fields, bringing us more accurate and smarter solutions.
However, to fully exploit the potential of large models faces great challenges during training and inference. The training of large models consumes a lot of computing resources and time, and the reasoning process also requires efficient computing power. In this context, Baidu Smart Cloud's large-scale model platform Wenxin Qianfan came into being, providing enterprise users with a full-process tool chain and a complete environment for large-scale model training and reasoning. By integrating high-performance computing facilities, optimized deep learning frameworks, and professional technical support, it helps companies train and deploy large models more efficiently, and use large models in the simplest and most efficient way.
In this article, we will deeply discuss the concept, application and challenges of large models, and introduce the large model platform of Baidu Smart Cloud in detail, so as to experience the excellent performance of large models through actual experience.

1. The development history of artificial intelligence

1.1 Traditional machine learning

Traditional machine learning refers to an earlier and classic type of machine learning method in the field of artificial intelligence. The characteristic of traditional machine learning is that it mainly relies on manually designed feature extraction and selection algorithms, including support vector machines, decision trees, random forests, naive Bayesian, logistic regression, K nearest neighbors, etc. These algorithms use statistics and optimization methods to extract and select appropriate features from input data and construct appropriate mathematical models to predict or classify data. .

The advantage of traditional machine learning is that their theoretical basis is relatively mature, the speed of training and reasoning is relatively fast, and it can be applied to various types of data. In addition, for some small-scale data sets, the effect of traditional machine learning methods is relatively good. However, traditional machine learning methods also have quite obvious limitations. For example, since traditional machine learning methods rely on manually selected features, it is difficult to capture complex nonlinear relationships in data; at the same time, these methods usually do not have adaptive learning capabilities and require human intervention to adjust the model.

1.2 Deep Learning

Deep learning ( Deep Learning, DL) realizes the learning and abstract representation of data through multi-level neural network models. Compared with traditional machine learning, deep learning does not need to manually design features, but automatically learns high-level feature representation of data through end-to-end training on large-scale data sets.
The core of deep learning is a deep neural network, which can be understood as a stacked structure composed of multiple neural network layers. Each layer contains a set of learnable parameters for mapping input data to a more abstract representation space. Deep learning optimizes these parameters through the iterative forward propagation and back propagation process, so that the network can learn and extract useful features from the data.

Deep learning has achieved major breakthroughs and applications in various fields, such as computer vision, natural language processing, speech recognition, and recommender systems. It can handle large-scale and high-dimensional data, has strong expression and generalization capabilities, and can effectively solve complex tasks such as image classification, target detection, semantic understanding, machine translation, etc. The classic deep neural network architecture example is as follows:

However, deep learning also faces some challenges. First, the training of deep learning models usually requires a large amount of labeled data and powerful computing resources. Second, problems such as overfitting and gradient disappearance still exist, which need to be dealt with by regularization, optimization algorithms, and parameter initialization. Finally, the classic deep learning model is task-oriented, and different tasks require separate design of model structure and training tuning, that is, the scalability or adaptability of downstream tasks of large models is better.

1.3 The Era of Large Models

With the wide application of ChatGPTsuch 文心一言large models in natural language processing and other fields, the concept of large models has begun to receive more and more attention. On the one hand, large models have achieved remarkable results in many fields, providing a strong support for the development of artificial intelligence technology. For example, in the field of natural language processing, large models can achieve more accurate and faster text generation, machine translation and other tasks; in the field of image recognition, large models can achieve higher accuracy tasks such as image classification and object detection. On the other hand, with the continuous advancement of hardware technology and the gradual popularization of computing resources, the application of large models has gradually expanded to more fields. With the increase of data sets and the complexity of model structure, researchers began to build larger and deeper neural network models to improve the model's ability to understand, express and generate.

In the era of large models, a series of deep learning models with huge parameters have hundreds of millions or even billions of parameters, such as BERT, GPT, , Transformeretc. The emergence of large models has made deep learning a milestone breakthrough in the fields of natural language processing, image recognition, and speech recognition.
The advantage of large models is that they can learn richer and more accurate representations from large-scale data, thereby improving the generalization and expression capabilities of the model. Large models have achieved impressive performance in some complex tasks. For example, in the field of natural language processing, using large models can achieve more accurate text classification, named entity recognition, semantic understanding, etc.

2. Wenxin Qianfan

Although large models have amazingly good performance, they also face some challenges and limitations. First of all, large models require a lot of computing resources and storage space, and have high requirements for hardware equipment. Secondly, the training time of large models is long, which takes a lot of time and cost. In addition, there are certain difficulties in the deployment and application of large models because they require high inference speed and memory footprint. Therefore, for ordinary users and enterprises, it is a natural choice to choose a cloud service platform to simplify these tasks, and 文心千帆as the first one-stop enterprise-level large model platform, it becomes the first choice for large model training, reasoning and deployment.
At present, Baidu Smart Cloud has opened the public version testing service of Wenxin Qianfan Large-scale Model Platform for enterprises and individual customers. The official application address is: https://cloud.baidu.com/survey/qianfan.html .

2.1 Introduction to Wenxin Qianfan

Wenxin Qianfan large-scale model platform is the official and only commercial service platform of Baidu Wenxin large-scale model. It is the world's first one-stop enterprise-level large-scale model platform created by Baidu Smart Cloud. Enterprises can not only directly call Wenxin Yiyan service on Wenxin Qianfan, but also provide enterprises with rich functions and flexible scalability, so that enterprises can customize and deploy large models suitable for them according to their own needs, so that enterprises can quickly use large models and make good use of large models.

Wenxin Qianfan large-scale model platform provides a convenient and easy-to-use tool chain, allowing enterprise users to quickly get started and use large-scale models. At the same time, it has full-process functions, covering the entire process from model training to final deployment, and provides enterprises with end-to-end solutions. It is one of the best choices for enterprises to embrace large-scale models.

2.2 Application Scenarios of Wenxin Qianfan

Wenxin Yiyan is a conversational language model that can respond to user instructions through natural language interaction and complete a series of tasks, including question and answer, text creation, and code error checking. Wenxin Qianfan's cloud service interface is Wenxin Yiyan's enterprise version cloud service, including a series of development and application tool chains such as large-scale model fine-tuning, which can efficiently access customer business systems, and support intelligent question and answer, content creation, product marketing and code writing and other rich scenarios:

• Customer support: answer common questions from users, provide product or service information, solve simple problems, provide basic technical support, etc.
• Personalized Assistant: According to the needs and preferences of users, provide them with personalized suggestions, recommendations and guidance, such as travel planning, movie recommendations, food recommendations, etc.
• Education and training: As part of the online learning platform, provide students with support such as answering questions, tutoring learning, and providing explanations of knowledge points
• Content generation: automatically generate text content based on user input, such as article drafting, story creation, short stories, etc.
• Emotional interaction: communicate emotionally with users, provide comfort, support and encouragement, similar to the role of emotional support robots
• Financial industry: Provide dialogue capabilities for digital employees, analysis of financial contract documents, double-record quality inspection semantic accuracy, and generation of marketing copy

2.3 Advantages of Wenxin Qianfan Platform

Large-scale models such as generative artificial intelligence have extremely high requirements for computing power. They must have large computing power, large models, and big data. Relying on Baidu's leading products in the four layers of chips, frameworks, models, and applications, Wenxin Qianfan, the large-scale model platform of Baidu Smart Cloud, provides comprehensive support and solutions.
In terms of knowledge enhancement, Baidu has built a knowledge map of about 550 billion facts, which integrates learning from massive data and large-scale knowledge. It can also directly call the knowledge map for knowledge reasoning, automatically build prompts, and efficiently meet user needs.
In terms of retrieval enhancement, Baidu has the largest Chinese search engine in the world. Baidu Search has developed a new generation of search architecture based on semantic understanding and matching. It deeply understands user needs and web content, performs semantic matching, and obtains more accurate search results. In turn, it provides high-accuracy and time-sensitive reference information for large models to better meet user needs.
In terms of dialogue enhancement, based on the accumulation of dialogue technology and applications, Wenxinyiyan has the ability of memory mechanism, context understanding and dialogue planning to achieve better dialogue coherence, rationality and logic.
In terms of model reasoning, the R&D team has significantly improved the performance of Wenxin large models in tasks such as logical reasoning, mathematical calculation and code generation through large-scale logical data construction, logical knowledge modeling, combination of coarse-grained and fine-grained semantic knowledge, and symbolic neural network technology.

3. The first experience of Wenxin Qianfan

Recently, I was invited to experience the Wenxin Qianfan large-scale model platform of Baidu Smart Cloud , and tested the model performance in multiple application scenarios.

3.1 Registration process

(1) Log in to the official application address: https://cloud.baidu.com/survey/qianfan.html , click Login Now, and choose the appropriate login method to log in to Baidu Smart Cloud:

(2) After logging in, real-name authentication is required first to obtain a complete large-scale model experience:

(3) Select account attribution, here use personal authentication:

(4) Select the authentication method. If you choose face authentication, fill in the relevant information, and use WeChat to scan the QR code to complete the authentication; if you choose bank card authentication, fill in the relevant information, and use your mobile phone number to complete SMS authentication:

(5) At the official application address: https://cloud.baidu.com/survey/qianfan.html , apply to experience Wenxin Qianfan service, fill in the relevant information, and submit the application:

(6) After the application is approved, click to use immediately:

(7) Before using it, you need to activate the payment on the console:

After successful payment and activation, you can use the Wenxin Qianfan large-scale model platform:

3.2 Create an application

(1) In order to call the Wenxin Qianfan service, you need to create an application in the Wenxin Qianfan console first:

(2) Name the application, select the required interface service, enter the application description, and click the Create Now button to complete the creation:

After the creation is complete, the created page will pop up:

(3) Under Select Application in the Parameter Configuration tab, select the application created in the previous step, and you can start to experience the excellent performance provided by Wenxin Qianfan:

3.3 Online Test

Next, online tests are mainly conducted from the aspects of semantic understanding, content creation, logical reasoning, programming ability and mathematical ability.

3.3.1 Semantic understanding

Semantic understanding refers to the in-depth analysis and interpretation of natural language to understand the meaning and context expressed in it.

Simple semantic understanding
Simple semantic understanding can judge whether the computer can accurately understand human language and make appropriate responses, allowing the model to answer "what are the applications of semantic understanding in the field of artificial intelligence?", Wenxinyiyan model concisely gives the application of semantic understanding in different fields:

Classical Chinese Comprehension Classical
Chinese is expressed in a concise and clear way, but also contains rich meanings and connotations. Whether or not classical Chinese can be well understood reflects a language model’s ability to understand complex language structures, ancient cultural backgrounds, and ideological systems. Next, let the Wenxin Yiyan model translate the fragments of "The Story of Yueyang Tower". You can see that Wenxin Yiyan can accurately and comprehensively interpret the language structure and ideological meaning of classical Chinese texts, which is only slightly different from the standard translation in the modification method:

Contextual understanding
Contextual understanding refers to the ability to accurately understand and infer the meaning of a text based on contextual information and context when reading or interpreting a text. Next, Wenxin Yiyan's contextual understanding ability was tested through multiple rounds of questions and answers. It can be seen that Wenxin Yiyan can accurately combine the context and specific context, and give correct and detailed answers:

3.3.2 Logical reasoning

Logical reasoning requires drawing new conclusions from known information, or judging the relationship between things through analysis and comparison.

Simple Reasoning
Let’s start with simple reasoning. Reasoning ability is one of the essential capabilities of large models. It can be seen that the Wenxin Yiyan model has clear logic. It first gives a clear answer, and then gives a logical reasoning process:

Logical pitfalls
Logical reasoning entails avoiding logical holes, ie erroneous inferences or unreasonable arguments. Next, increase the complexity of the reasoning and use a common "teaching" with logical questions as a test. You can see that Wen Xin Yi Yan can accurately discover the logical trap, that is, there is no necessary causal problem between the premises of the reasoning, ignoring the influence of other factors:

Logical interference
Logical interference refers to the appearance of some interference or erroneous factors during logical reasoning, which lead to deviations or inaccurate conclusions in the reasoning process. In this test problem, a lot of irrelevant information is given to interfere with the model’s understanding of the nature of the problem, but it can be seen that Wen Xinyiyan has a clear understanding of the requirements and conditions of the problem, avoiding the interference of irrelevant information, deriving correct conclusions, and avoiding wrong inferences caused by logical interference:

Multi-chain reasoning
Multi-chain reasoning refers to the reasoning process in which the final conclusion is reached by connecting multiple intermediate steps or logical chains in logical reasoning, which is used for the analysis and solution of complex problems. Next, test Wenxinyiyan's multi-chain reasoning performance through a classic multi-chain logic problem. It can be seen that the model can obtain the correct solution through logical connection and derivation, has complex logical reasoning capabilities, and can provide comprehensive and accurate answers:

3.3.3 Content Creation

Content creation is one of the common application areas for large models. Large models need to have strong language generation capabilities and be able to generate smooth, coherent and creative texts based on input information.

Article Writing
Article writing is a common form of writing that can be used for academic, exam, competition or other purposes. After giving a clear article theme, main points and writing purpose, the big model should have a clear understanding of its meaning and requirements, and ensure the overall quality and integrity of the article. Through the writing task of "writing a composition describing apples, expressing the perception and taste of apples" to test Wenxinyiyan's article writing ability, it can be seen that Wenxinyiyan can use clear, accurate and fluent language description language, and adopts the structure of total score to organize key points, the transition between paragraphs is natural and smooth, and the structure of the whole article is compact and clear:

Film review writing
Film review writing aims to analyze, comment and evaluate films or film and television works. By "writing a film review about The Amazing Spider-Man 2" to test Wenxin Yiyan's film review writing performance, we can see that the model can not only fully introduce the theme, genre and background information of the movie, but also understand that this movie is a sequel to The Amazing Spider-Man, with a good structure and clear logic, covering the main content introduction of the film, film production evaluation, actor performance evaluation, story and plot evaluation, etc., so that readers can have a clear and comprehensive understanding of this movie:

Soft text writing
Soft text writing is a form of copywriting to promote and promote products, services or brands. Soft text writing is an important manifestation of the ability to create large models. Next, let Wenxinyiyan write an advertorial based on the requirement of "writing an advertisement about TV, integrating humanistic care, and needing euphemistic recommendation". You can see that Wenxinyiyan can adjust the language style and expression method according to the needs and requirements of users, and use extensive knowledge, language comprehension ability and rich rhetorical techniques to generate creative copywriting content, making the advertorial more attractive and convincing:

3.3.4 Mathematics ability

Mathematics ability includes basic mathematical calculations, such as addition, subtraction, multiplication and division, square root, exponentiation and other operations, as well as complex mathematical knowledge and concepts, such as algebra, geometry, probability and statistics, etc. Large models need to be able to process and understand complex mathematical problems and generate relevant answers or solutions.

Simple Math Problems
Take the simple encounter problem as an example to calculate whether two objects will meet based on information such as their speed, direction, and initial position. It can be seen that Wenxin Yiyan can quickly generate accurate calculation results, and contains reasonable calculation processes, thus saving time and effort for manual calculations:

Complex math problems
Complex math problems usually need to consider more conditions. In Mathematical Olympiad, you often encounter the problem of cattle grazing. Not only the speed of cattle grazing, but also the growth speed of grass must be considered. Wenxin Yiyan can not only give correct results, but also give detailed intermediate steps and problem-solving ideas, which can help users better understand mathematical knowledge:

3.3.5 Programming ability

Programming ability includes the ability to solve tasks such as code completion, document generation, question answering, and code conversion, and is an important aspect of measuring large model-assisted programming.

Simple Questions
First, take "implementing binary sorting method with Python" as an example to test the performance of the model on simple programming problems. You can see that if you ask questions about programming to the Wenxin Yiyan model, it can give answers and provide comments on the core ideas of the algorithm, which can help users better understand and solve problems:

Complex Problems
We have seen that the Wenxin Yiyan model can solve simple programming problems very well. Next, we will further test the complex problem-solving ability of the model. Taking the classic recursive problem-the Tower of Hanoi as an example, the Wenxin Yiyan model first interprets the topic statement, then gives the correct code, and finally gives a test code example, which can give users inspiration and guidance when solving problems:

Finding bugs
and modifying bugs is the norm for most programmers in their daily work. Therefore, if you can use Wenxinyiyan to find and modify bugs in the code, it will greatly improve the programmer's work efficiency. Next, take the following code as an example to test the bug fix performance of Wenxinyiyan:

l = [1, 2, 3, 4, 5]
for i in l:
    l.append(i)

In the figure below, you can see that the Wenxinyiyan large-scale model can be used to quickly debug and verify the code, and try to give the correct answer, quickly verify the code logic and eliminate potential errors:

Code Comments
Code comments are text added to the program code. Its function is to explain the function, logic or design intent of the code to other developers or yourself. Good code comments can greatly improve the efficiency of teamwork. Next, we let the Wenxinyiyan model explain the following code:

void QuickSort(int *arr,int low,int high){
    
    
		if(low<high){
    
    //不加这个条件会死循环 
		int i=low;
		int j=high;
		int key=arr[low];
		while(i<j){
    
    
			while(i<j&&arr[j]>=key)
				j--;
			if(i<j) 
				arr[i++]=arr[j];//先赋值在自增 
			while(i<j&&arr[i]<key)
				i++;
			if(i<j)
				arr[j--]=arr[i];//先赋值在自增				
		}
		arr[i]=key;
		QuickSort(arr,low,i-1);//给左边排序 
		QuickSort(arr,i+1,high);//给右边排序 
	}
}

It can be seen that Wenxin Yiyan can correctly provide text explanations for codes to help users understand the functions and logic of the codes. At the same time, it also gives a flow description of the algorithm, which can explain complex algorithms, business logic or design ideas, making the code easier to read and understand:

In addition to the online experience, the Wenxin Qianfan large model platform also provides full-process services from data services to model training. Users can choose a training mode suitable for their task scenarios according to their business needs and conduct parameter adjustment training to achieve ideal model effects. Next, Wenxin Qianfan provides data services and large model training functions.

3.4 Data Services

Wenxin Qianfan platform supports unified management of training model data sets, and supports operations such as independent version iteration, data viewing, import, export, and deletion. Select Dataset Management in the function column on the left side of the console to enter the main task interface of Dataset Management, where texts of different annotation types are used for different model training methods.

Dataset annotation types are divided into the following two types:

• Text conversation:
  • Non-sorting template: single or multiple rounds of text dialogue data, one-to-one correspondence between a single question and answer, as the training input for subsequent SFT, prompt tuning, and delta tuning
  • Including sorting template: single or multiple rounds of text dialogue data, a single question corresponds to multiple answers, multiple answers need to be sorted, as the training input of the subsequent reward model
• query question set: single or multiple rounds of question corpus data, used as training input for subsequent RLHF

After creating a dataset, find the dataset on the dataset management page, and click the "Import" button under the operation column on the right to enter the import data page:

On the Import Data page, you can upload text data using a number of different options:

After the data is imported, you can choose to specify an unpublished dataset version for labeling in the dataset list, or specify an unreleased dataset version for labeling in the labeling interface:

In order to improve the data quality, the platform supports the cleaning of text datasets. Select data cleaning in the function column data processing on the left side of the Wenxin Qianfan large model console, and enter the main task interface of data cleaning:

3.5 Large model tuning

Large-scale model tuning, also known as large-scale model fine-tuning ( Fine-Tuning), refers to the process of further training a model that has been pre-trained on large-scale data using a relatively small task-specific data set. Developers can choose a training mode suitable for their own task scenarios and perform parameter adjustment training to achieve the ideal model effect.
In order to meet the individual needs of users, Wenxin Qianfan provides a large-scale model tuning function. Select large-scale model tuning in the function column on the left side of the Wenxin Qianfan large-scale model console to enter the main task interface of large-scale model tuning:

On the large model tuning task interface, select the "Create tuning task" button, fill in the task name, select the industry and application scenario, and click Create to create the tuning task:

When creating a task, select "Create and Train", or in the large model tuning task list, select the "New Run" button of the specified task to enter the task configuration page for model training, configure the parameters for training the large model, adjust the basic configuration, and click "Confirm" to start the task of model training:

3.6 Tips for using Prompt

PromptIt is an input text provided to the model, usually a question, a sentence or a piece of text, which is used to guide the model to generate relevant answers or complete specific tasks. PromptThe accuracy, clarity, and relevance of the model are very important for the answers generated by the model. Reasonable selection and design of prompts can obtain more accurate and useful responses.
If Promptnot designed properly, it can lead to poor quality model generation and even inaccurate or wrong text. Therefore, Promptit is very important to optimize and improve the design and usage of . Through optimization and improved Promptdesign methods, the pertinence and accuracy of the model can be improved, resulting in higher quality text generation.
For example using Prompt: "Tell me about animals.", because it is too vague to provide specific information or guidance, the model may generate irrelevant or ambiguous responses. When using Prompt: "Please introduce the characteristics and living habits of penguins.", because it is specific and clear, the required information is provided, and the model can generate relevant and accurate Promptanswers based on this:

Good PromptIt is clear, specific and detailed, and covers key information and requirements to help the model generate accurate and relevant answers. According to specific tasks or needs, it can be flexibly adjusted and optimized to obtain the best results Prompt.

4. User experience and development prospect

As a developer, I am very honored to be able to experience and explore the potential and possibilities brought by Wenxin Qianfan. From the perspective of user experience, the Wenxin Yiyan model has demonstrated impressive natural language generation capabilities. Whether faced with common questions or more challenging tasks, it answers with fluidity, accuracy and creativity. During the chat process, it can understand the context and provide targeted responses, making the conversation more coherent and effective. It can also provide practical suggestions, explain complex concepts and help solve problems, enabling users to obtain a personalized and customized interactive experience. However, the answer of the Wenxinyiyan large model may be affected by bias and errors in the training data. In order to avoid the impact of these misleading information on the performance of the model, on the one hand, we can improve the quality, and on the other hand, we can also fine-tune the model to make it more in line with business expectations Prompt.
With more user usage and feedback, researchers and developers will be able to collect a large amount of data and feedback to improve and optimize the performance and performance of the model, which will help improve the accuracy, reliability and creativity of Wenxin Qianfan, making it gradually become a powerful and reliable smart assistant. Similarly, with the deepening of research, Wenxin Qianfan large-scale model platform will be applied to more practical scenarios and fields, for example, in the fields of customer service, education and training, medical consultation, etc., to provide people with personalized support and advice, and improve user experience and satisfaction. With the advancement of technology and the expansion of applications, Wenxin Qianfan large-scale model platform is expected to bring innovation and change to all walks of life.

summary

The advantage of large models is that they can handle more complex and higher-level abstraction tasks, improve the performance and effect of natural language processing and machine learning applications, and have strong context understanding capabilities, language generation capabilities, and multi-field adaptability. However, in order to give full play to the potential of large models, there are also a series of problems such as data processing and labeling, computing resource requirements, training time and cost. In order to solve this problem, Baidu Smart Cloud highlights the large model platform Wenxin Qianfan , which provides enterprise users with a full-process tool chain and a complete environment for large model training and reasoning. By providing rich model selection, powerful computing resources and professional technical support, users can more efficiently apply large models to solve complex tasks and problems. In this article, I deeply experienced the superior performance of Wenxin Qianfan large model platform, and demonstrated its excellent performance in semantic understanding, content creation, logical reasoning, programming ability and mathematical ability. Wenxin Qianfan application experience address: https://cloud.baidu.com/survey/qianfan.html , come and experience it together!