AI Platform Scenario-specific Model Tuning Service

Service Overview

The AI Platform Development and Implementation Service targets customers across various industries with specific business scenarios, aiming to address these needs through AI technologies. Leveraging the AI platform, Huawei Cloud offers professional scenario-specific model tuning service along with incremental model training, distillation, and fine-tuning support to tackle critical challenges like complex algorithm development and intricate model tuning. By purchasing this professional service, you can establish a comprehensive AI solution with the help of Huawei, swiftly achieving end-to-end AI scenario-specific modeling, tuning, and ecosystem transition tailored to your specific needs.

Service Content

• AI Platform Scenario-specific Model Tuning Service

  Specifications	Service Content	Application Scenarios
  AI Platform Scenario-specific Model Tuning Service - Basic	Requirement survey: Implement, launch, and verify the solution based on the technical solution. Solution design: Conduct a business scenario survey, identify the customer's business pain points, and produce a requirement analysis report based on ModelArts. Solution implementation: Based on customer requirements and pain points, optimize algorithm metrics and verify the optimization using offline data.	The customer needs Huawei's guidance to optimize the metrics or accuracy of the delivered solution. The optimization must be verified with offline data.
  AI Platform Scenario-specific Model Tuning Service - Standard	Requirement survey: Implement, launch, and verify the solution based on the technical solution. Solution design: Conduct a business scenario survey, identify the customer's business pain points, and produce a requirement analysis report based on ModelArts. Solution implementation: Optimize algorithm metrics based on customer requirements and pain points, and test the optimization in real businesses.	The customer needs Huawei's guidance to optimize the metrics or accuracy of Huawei Cloud AI assets. The optimization must be verified with real business data.
  AI Platform Scenario-specific Model Tuning Service - Professional	Requirement survey: Implement, launch, and verify the solution based on the technical solution. Solution design: Conduct a business scenario survey, identify the customer's business pain points, and produce a requirement analysis report based on ModelArts. Solution implementation: Optimize algorithm metrics based on customer requirements and pain points, test the optimization in real businesses, and develop new use cases.	The customer needs Huawei's guidance to optimize the metrics or precision of delivered solutions for applying them to other scenarios. The optimization must be verified with real business data and new use case needs to be provided.
  AI Platform Scenario-specific Model Tuning Service - Platinum	Requirement survey: Implement, launch, and verify the solution based on the technical solution. Solution design: Conduct a business scenario survey, identify the customer's business pain points, and produce a requirement analysis report based on ModelArts. Solution implementation: Optimize algorithm metrics based on customer requirements and pain points, test the optimization in real businesses, and develop new use cases.	The customer needs Huawei's guidance to optimize the metrics or precision of delivered AI assets or solutions for applying them to other scenarios. The optimization must be verified with real business data and new use case needs to be provided.
  Model Distillation Service - Standard	Using DeepSeek models as teachers, provide data distillation services for foundational model capabilities like Q&A, copywriting, and reading comprehension. Each set includes 2,000 pieces of data from industry instruction datasets.	Suitable for customers seeking distilled datasets with chain-of-thought capabilities and strong logical reasoning abilities
  Model Distillation Service - Professional	Using DeepSeek models as teachers, provide data distillation services for foundational model capabilities like Q&A, copywriting, and reading comprehension. Each set includes 5,000 pieces of data from industry instruction datasets. The service also offers model distillation for written data.	Suitable for customers seeking domain-specific large AI models with chain-of-thought capabilities and strong logical reasoning abilities
  Model Distillation Service - Platinum	Using DeepSeek models as teachers, provide data distillation services for foundational model capabilities like Q&A, copywriting, and reading comprehension. Each set includes 5,000 pieces of data from industry instruction datasets. The service also offers model distillation and reinforcement learning for written data.	Suitable for customers seeking domain-specific large AI models with chain-of-thought capabilities and very strong logical reasoning abilities
  Incremental Model Training Service - Standard	Leveraging the DeepSeek model, performs incremental model pre-training using extensive industry pre-training data and conducts SFT fine-tuning using a substantial no-reasoning industry instruction dataset. These two training phases assist customers in developing industry-specific large AI models.	Suitable for customers seeking industry-specific large AI models with moderate logical reasoning abilities but no chain-of-thought capabilities
  Incremental Model Training Service - Professional	Leveraging the DeepSeek model, performs incremental model pre-training using extensive industry pre-training data, conducts SFT fine-tuning using a substantial no-reasoning industry instruction dataset, and implements reinforcement learning. These three training phases assist customers in developing industry-specific large AI models.	Suitable for customers seeking industry-specific large AI models with strong logical reasoning abilities but no chain-of-thought capabilities
  Incremental Model Training Service - Platinum	Leveraging the DeepSeek model, performs incremental model pre-training using extensive industry pre-training data, conducts SFT fine-tuning using a substantial reasoning industry instruction dataset, and implements reinforcement learning. These three training phases assist customers in developing industry-specific large AI models.	Suitable for customers seeking industry-specific large AI models with chain-of-thought capabilities and very strong logical reasoning abilities
  Model Fine-Tuning Service - Standard	Service content: Fine-tune data assets based on the algorithm models included in the Huawei open-source adaptation list, foundational model capabilities (Q&A, copywriting, and reading comprehension), and customers' industry instruction data. The service also offers fine-tuning for written data. Deliverables: 2,000 Q&A pairs for model fine-tuning. Acceptance criteria: Based on training data writing standards recognized by both parties, the customer provides a sealed acceptance report after acceptance or performs online acceptance.	Suitable for customers with limited industry instruction knowledge who want to quickly create scenario-specific models
  Model Fine-Tuning Service - Professional	Service content: Fine-tune data assets based on the algorithm models included in the Huawei open-source adaptation list, foundational model capabilities (Q&A, copywriting, and reading comprehension), and customers' industry instruction data. The service also offers fine-tuning for written data. Deliverables: 5,000 Q&A pairs for model fine-tuning. Acceptance criteria: Based on training data writing standards recognized by both parties, the customer provides a sealed acceptance report after acceptance or performs online acceptance.	Suitable for customers with rich industry instruction knowledge who want to quickly create scenario-specific models
  Model Fine-Tuning Service - Platinum	Service content: Fine-tune data assets based on the algorithm models included in the Huawei open-source adaptation list, foundational model capabilities (Q&A, copywriting, and reading comprehension), and customers' industry instruction data. The service also offers fine-tuning and reinforcement learning for written data. Deliverables: 5,000 data records for model fine-tuning Q&A and reinforcement learning. Acceptance criteria: Based on training data writing standards recognized by both parties, the customer provides a sealed acceptance report after acceptance or performs online acceptance.	Suitable for customers with rich industry instruction knowledge who want to quickly create scenario-specific models to improve decision-making in complex tasks

Prerequisites

• Customers shall submit service requests at least 10 working days in advance to allow Huawei Cloud to assess their needs and coordinate AI experts.
• Both parties have agreed on the service objectives and have signed a contract.

Service Scope

1. Items Coverage

   Huawei offers AI Platform Development and Implementation Service tailored for AI application development and research across diverse industries. Catering to diverse customer needs at different stages, the service includes scenario-specific model tuning, incremental training, fine-tuning, and distillation. This helps businesses seamlessly incorporate AI capabilities into their service applications, covering areas like computer vision, machine learning, NLP, and decision optimization. Additionally, Huawei provides services for planning and design as well as model tuning for industries such as industrial quality inspection, sound quality inspection, process optimization, retail identification, and intelligent scheduling. Additionally, the service offers comprehensive development support, including application & model development sample demonstrations and model & application migration.

2. Items Not Covered

   Huawei AI engineers will only deliver the services specified in this statement of work (SOW). Any services outside of this scope will incur additional charges. These excluded services include, but are not limited to:

   • Working beyond regular working days and hours as required by the customer (note that overtime pay and allowances apply).
   • Purchasing products from other companies or individuals.
   • Performing development and maintenance tasks associated with customer services that are beyond the defined scope in the project solution.
3. Service Regions

   Outside China (with product globalization strategy).

Service Process

• Service process of AI Platform Scenario-specific Model Tuning Service

  Service phase	Description
  Requirement survey and evaluation	Specify the requirement scope for development, deployment, and tuning services and assess their feasibility.
  Solution design	Design development, deployment, and tuning service solutions.
  Solution implementation	Implement scenario-based modeling, deployment, and tuning solutions.
  Service acceptance	The customer verifies the items in the deliverables and signs the AI Platform Development and Implementation Service Acceptance Report.
  Project handover	Hand over related deliverables to the customer to signify project completion.

  

• Service process of Model Distillation Service

  Service phase	Description
  Scenario analysis	Analyze the customer's business scenarios, data status, and model foundations, clarify distillation objectives and requirements, and define model input/output, operating environments, and performance requirements.
  Distillation solution design	Specify the distillation objectives and requirements, determine model input and output, operating environments, and performance requirements, select a suitable student model, and develop a distillation policy incorporating data fine-tuning and reinforcement learning.
  Data distillation implementation	Based on the distillation scheme, select domain-specific and general datasets, utilize industry data to create seed data, employ the teacher model to generate and refine teacher data, and appropriately balance general data with teacher data to prepare high-quality data for model distillation.
  Distillation solution implementation	Utilize techniques like SFT, LoRA, and RFT for model fine-tuning, employ reinforcement learning methods including DPO, PPO, and GPRO to enhance model performance, leverage approaches such as mixed-precision training acceleration and model quantization to boost operational efficiency, thereby completing model tuning and training.
  Distillation scenario verification	Perform an end-to-end test on the distilled student model, comparing its performance, throughput, latency, and memory usage with the pre-distillation model. Then deploy it in real-world business scenarios to validate the alignment between model performance and business metrics, ensuring the distillation outcome meets expectations.
  Service acceptance	The customer verifies the items in the deliverables and signs the Model Distillation Service Acceptance Report.

  

• Service process of Incremental Model Training Service

  Service Phase	Description
  Incremental training solution design	Specify the rationale and feasibility of incremental training, and detail the training data, model training, and model evaluation solutions.
  Data solution implementation	Acquire, process, assess, and match data while ensuring its security.
  Model solution implementation	Prepare the environment, initiate the incremental training workflow, and offer precision and performance tuning services.
  Incremental training scenario verification	Verify the precision performance, evaluate the model, and assess and validate the business impact.
  Acceptance	The customer verifies the items in the deliverables and signs the DeepSeek Incremental Model Training Service Acceptance Report.
  Handover	Hand over related deliverables to the customer to signify project completion.
  Delivery confirmation	Confirm the accuracy and completeness of deliverables and confirm the completion of the project.

  

• Model fine-tuning process

  Service Phase	Description
  Scenario-specific solution design	Select a proper fine-tuning route based on customer service requirements and data status, such as supervised fine-tuning (SFT, LoRA, and QLoRA) training scenarios.
  Data solution design	Design data solutions for different fine-tuning routes, including data collection, labeling, cleansing, and review.
  Model solution design	Configure model hyperparameters based on customer requirements and hardware resources, such as the learning rate, batch size, and number of training rounds.
  Evaluation solution design	Construct evaluation datasets, including representative samples extracted from customers' actual operations. Also, define model evaluation criteria to assess the model's effectiveness.
  Data solution implementation	Data collection: Collect relevant data from customers' business systems, databases, log files, and other channels. Data labeling: Label the collected data according to the designed instruction format and task requirements. Data cleansing: Eliminate duplicate, erroneous, and irrelevant data records, and address issues such as missing values. Data distillation: Utilize industry-specific prompt data to extract answers from models and construct question-and-answer pairs into instruction datasets. Data ratio: Determine the ratio of industry data to general data based on the characteristics of customers' industry data and model training requirements.
  Model development environment preparation	Ensure that appropriate hardware and software environments are available. Use the AI Cloud Service platform to acquire compute resources. Install essential dependency libraries and tools.
  Model fine-tuning training	Use fine-tuning training methods such as SFT, LoRA, and QLoRA to train the model based on the designed fine-tuning solution.
  Model effectiveness tuning	During the training, closely monitor the changes in metrics such as the loss value and accuracy of the model, and promptly adjust the training parameters and policies.
  Fine-tuning scenario verification	Precision and performance verification: Verify model performance indicators like model training throughput and training loss convergence. Model evaluation and verification: Objectively assess the model's scoring results based on the constructed evaluation dataset. Service effectiveness verification: Subjectively evaluate the model's generated results or predictions in detail against criteria such as accuracy, completeness, relevance, efficacy, fluency, and within specific business scenarios.

  

• Model fine-tuning process - model reinforcement learning process

  Service Phase	Description
  Scenario-specific solution design	Select an appropriate reinforcement learning approach, such as PPO or GRPO, based on the performance of the fine-tuned model and the state of the data.
  Data solution design	Design reinforcement learning data collection policies for enhanced pathways to determine how to collect and construct datasets that accurately mirror human preferences.
  Model solution design	Reward function/Reward model design: Assess the quality of responses generated by the model, aid in annotating reinforcement learning data, and deliver precise reward signals for model training. Configure model hyperparameters based on customer requirements and hardware resources, such as the learning rate, batch size, and number of training rounds.
  Evaluation solution design	Construct evaluation datasets, including representative samples extracted from customers' actual operations. Also, define model evaluation criteria to assess the model's effectiveness.
  Data solution implementation	Preference data labeling: Organize manual or tool-based preference labeling on data according to the designed reward function/model, giving likes to answers that meet expectations and dislikes otherwise, actively guiding the model with bias.
  Model development environment preparation	Ensure that appropriate hardware and software environments are available. Use the AI Cloud Service platform to acquire compute resources. Install essential dependency libraries and tools.
  Fine-tuning solution implementation	Use reinforcement learning methods such as PPO and GRPO to further optimize the fine-tuned model based on the designed reward model.
  Fine-tuning scenario verification	Precision and performance verification: Verify model performance indicators like model training throughput and training loss convergence. Model evaluation and verification: Objectively assess the model's scoring results based on the constructed evaluation dataset. Service effectiveness verification: Subjectively evaluate the model's generated results or predictions in detail against criteria such as accuracy, completeness, relevance, efficacy, fluency, and within specific business scenarios.

  

Service Deliverables

• AI Platform Scenario-specific Model Tuning Service

  Service	Deliverables
  AI Platform Scenario-specific Model Tuning Service - Basic	Verification Results of ModelArts Scenario-specific Model Tuning
  AI Platform Scenario-specific Model Tuning Service - Standard
  AI Platform Scenario-specific Model Tuning Service - Professional
  AI Platform Scenario-specific Model Tuning Service - Platinum
  Model Distillation Service - Standard	AI Platform Scenario-specific Model Tuning Service - Model Distillation Service Development Report
  Model Distillation Service - Professional
  Model Distillation Service - Platinum
  Incremental Model Training Service - Standard	AI Platform Scenario-specific Model Tuning Service - Model Incremental Training Service Development Report
  Incremental Model Training Service - Professional
  Incremental Model Training Service - Platinum
  Model Fine-Tuning Service - Standard	AI Platform Scenario-specific Model Tuning Service - Model Fine-Tuning Service Development Report
  Model Fine-Tuning Service - Professional
  Model Fine-Tuning Service - Platinum

Responsibility Matrix

1. Common Responsibilities
   • Negotiate and confirm specific requirements and objectives.
   • Negotiate and confirm project management plans.
   • Negotiate, confirm, and review solutions.
   • Sign the contract.
2. Huawei Responsibilities
   • Huawei specifies a project owner. Customers should be notified of any personnel changes three working days in advance until the project is accepted.
   • Huawei can only use the data authorized by the customer.
   • Huawei proposes the guidance plan and quotation based on the selected services for the customer's review before the consultation.
   • Huawei provides technical guidance for the customer according to the confirmed plan during the consultation period.
   • Huawei provides the deliverable list based on the selected consultation service items after the consultation.
   • Huawei receives the service request from the customer and assigns ModelArts experts to discuss the details with the customer.
3. Customer Responsibilities
   • The customer provides detailed and accurate requirements and scenarios.
   • The customer provides necessary items for project implementation, such as training data.
   • The customer reviews and confirms the guidance plan and deliverables provided by Huawei.
   • The customer accepts the project.
4. Responsibility Details
   • AI Platform Scenario-specific Model Tuning Service

     The following table provides an example responsibility matrix and can be modified as needed. R = responsible party S = supporting party Note: If Huawei offers technical support for a service, the customer is responsible for its implementation. Given the specificity of performance optimization, the sequence of designing and implementing the optimization plan may be adjusted according to specific project circumstances.
     No.	Process	Task	Huawei	Customer
     1	Requirement proposal	Explain business scenarios and clarify optimization requirements.	S	R
     2	Requirement understanding	Understand the requirements and define key issues.	R	S
     3	Requirement confirmation	Requirement confirmation and acceptance item confirmation (key metrics)	S	R
     4	Solution design	Optimization solution design	R	S
     5	Solution report	Optimization solution report	R	R
     6	Solution confirmation	Confirm the optimization solution.	S	R
     7	Solution implementation	Implement the optimization solution.	R	S
     8	Solution acceptance	Accept the optimized performance.	S	R
     9	Project handover	Sort out and hand over deliverables.	R	S
     10	Delivery confirmation	Confirm the accuracy and completeness of deliverables and confirm the completion of the project.	S	R

   • Model Fine-Tuning Service

     The following table provides an example responsibility matrix and can be modified as needed. R = responsible party S = supporting party Note: If Huawei offers technical support for a service, the customer is responsible for its implementation.
     No.	Process	Task	Huawei	Customer
     1	Fine-tuning solution design	Design the scenario solution, data solution, model solution, and evaluation solution.	R	S
     2	Data solution implementation	Collect data, label data, cleanse data, distill data, proportion data, and label preference data.	R	R
     3	Model solution implementation	Prepare the environment, implement model fine-tuning, conduct reinforcement learning, and optimize model performance.	R	S
     4	Fine-tuning scenario verification	Verify the precision performance, evaluate the model, and verify the service effectiveness.	R	S
     5	Service acceptance	The customer signs to confirm the acceptance.	S	R
     6	Project handover	Sort out and hand over deliverables.	R	R
     7	Delivery confirmation	Customer acceptance	S	R

   • Incremental Model Training Service

     The following table provides an example responsibility matrix and can be modified as needed. R = responsible party S = supporting party Note: If Huawei offers technical support for a service, the customer is responsible for its implementation.
     No.	Process	Task	Huawei	Customer
     1	Incremental training solution design	Analyze scenarios and design data solutions, model solutions, and model evaluation solutions.	R	S
     2	Data solution implementation	Extract and cleanse data, label data, evaluate data, combine data, and ensure data security.	R	R
     3	Model solution implementation	Prepare the environment, implement incremental model training, and tune the model.	R	S
     4	Incremental training scenario verification	Verify the precision performance, evaluate the model, and verify the service effectiveness.	R	S
     5	Service acceptance	The customer signs to confirm the acceptance.	S	R
     6	Project handover	Sort out and hand over deliverables.	R	R
     7	Delivery confirmation	Customer acceptance	S	R

   • Model Distillation Service

     The following table provides an example responsibility matrix and can be modified as needed. R = responsible party S = supporting party Note: If Huawei offers technical support for a service, the customer is responsible for its implementation.
     No.	Process	Task	Huawei	Customer
     1	Distillation solution design	Analyze service scenarios and select student models and training policies.	R	S
     2	Data distillation implementation	Select datasets, generate seed data and teacher data, and combine data.	S	R
     3	Distillation solution implementation	Fine-tune and perform reinforcement learning on student models.	R	S
     4	Distillation scenario verification	Verify the performance and service metrics of the distilled model.	R	S
     5	Solution acceptance	Accept the model functions and performance.	S	R
     6	Project handover	Sort out and hand over deliverables.	R	S
     7	Delivery confirmation	Confirm the accuracy and completeness of deliverables and confirm the completion of the project.	S	R

Acceptance Criteria

Huawei submits the standard deliverables described in section "Service Deliverables" based on each service sub-item. If customers accept the deliverables, they need to click the acceptance link on the Huawei Cloud console or sign and seal the AI Platform Development and Implementation Service Acceptance Report.