International LLM Technical Services

1.1 Service description

LLM consulting service:

Scenario-based consulting for intelligent computing needs, covering the Qwen series of LLM products. Includes prompt optimization and model evaluation support.

Agent building technical service:

Agent building based on your business scenarios, including scenario analysis, architecture design, engineering workflow implementation, deployment, and performance monitoring and optimization.

Model fine-tuning and training service:

End-to-end services covering data preparation, model selection, training, and tuning for domain knowledge adaptation and model customization. Supported methods include SFT, LoRA, and DPO.

Supports CPT fine-tuning and training, with services for data cleaning, annotation, training monitoring, model validation, and deployment.

2.1 Scope of the LLM consulting service

The service scope includes:

1) Scenario and requirements research: We conduct interviews to identify business pain points and LLM application needs. We also research your existing resources, including knowledge bases, data assets, and API operations.

2) Current technology research: We analyze your computing resources, network environment, and system architecture. We also determine the integration capabilities and interface specifications for your knowledge sources, databases, and application systems.

3) Agent orchestration plan design: We design the agent workflow, nodes, model selection, and evaluation methods and standards.

4) Contextual engineering optimization plan: We provide in-depth prompt optimization services tailored to your scenarios.

5) Project process management and consulting plan development support.

The service scope does not include the following:

1) Engineering development related to agents.

2) Code development or feature modification for your systems.

3) Daily operations and maintenance (O&M) services for third-party software, such as installation, testing, troubleshooting, or optimization.

4) IT system architecture transformation or general digital transformation consulting not related to LLMs.

2.2 Scope of the agent building technical service

The service scope includes:

1) Research and analysis of your LLM-related application scenario requirements and technologies.

2) Agent architecture design, orchestration, and implementation.

3) Contextual engineering design and implementation.

4) Agent performance evaluation and optimization. This does not include model fine-tuning or training.

5) Integration of tools or Skills.

6) Project process management and implementation plan development support.

7) Current agent building scenarios include the following:

The service scope does not include the following:

1) Code development or feature modification for your systems.

2) Frontend application development or feature modification for the agent.

3) Daily O&M services for third-party software, such as installation, testing, troubleshooting, or optimization.

4) IT system architecture transformation or general digital transformation consulting not related to LLMs.

2.3 Model fine-tuning and training service

The service scope includes:

1) Research on your requirements for fine-tuning or training scenarios.

2) Design of the fine-tuning or training plan.

3) Dataset preparation, including training and evaluation data.

4) Performance evaluation and optimization.

5) Model deployment and publishing support.

6) Project process management and implementation plan development support.

The service scope does not include the following:

1) Code development or feature modification for your systems.

2) Installation, testing, troubleshooting, or optimization of third-party closed-source software.

3) Training and inference frameworks other than Alibaba Cloud products, or any framework-level modifications for either Alibaba Cloud or third-party frameworks.

3.1 Prerequisites for the AI consulting service

1) Service request timeline: You must request the service at least 15 calendar days in advance.

2) You must provide a business scenario description, existing system architecture documents, and a manifest of your knowledge bases and data assets.

3) You must purchase and use Alibaba Cloud Qwen series products. At a minimum, this includes the Qwen model.

4) Environment and permissions: You must provide access permissions to a non-production environment, a remote access channel, and the necessary data and code resources.

5) Project cooperation: You must designate a project manager or technical manager with decision-making authority as the key contact. You must also confirm the implementation plan in writing within 5 business days.

3.2 Prerequisites for the agent building technical service

1) Service request timeline: You must request the service at least 15 calendar days in advance.

2) Environment and permissions: You must provide access permissions to a non-production environment, a remote access channel, and the necessary data and code resources.

3) You must purchase Alibaba Cloud computing resources or use the Alibaba Cloud Model as a Service (MaaS) platform service.

4) Project cooperation: You must designate a project manager or technical manager with decision-making authority as the key contact. You must also confirm the implementation plan in writing within 5 business days.

3.3 Prerequisites for the model training service

1) Service request timeline: You must request the service at least 15 calendar days in advance. For scenarios that involve large-scale data or computing resource migration, you must request the service 30 days in advance.

2) You must provide the necessary non-production environment, data samples, API operation documents, and remote access permissions.

3) You must provide the training dataset, computing resources, clear task objectives, and evaluation standards.

4) Project cooperation: You must designate a project manager or technical manager with decision-making authority as the key contact. You must also confirm the implementation plan in writing within 5 business days.

5) You must purchase Alibaba Cloud computing resources or use the Alibaba Cloud MaaS or Intelligent Computing Platform service.

6) Alibaba Cloud does not design reinforcement learning reward functions. You must provide the reward function.

4.1 Roles and responsibilities

4.2  SLA

Alibaba Cloud will provide an LLM service technical manager.

Alibaba Cloud will provide an "LLM Technical Service Work Plan" and an "LLM Technical Service Acceptance Report".

5.1 AI consulting service items

Each package provides the following research and consulting services:

1) Provides research and analysis services on Large Language Model (LLM) requirements and scenarios, including the following:

a) Confirm the customer's business pain points and Large Language Model (LLM) application requirements through an interview.

b) You can assess the status of active resources, such as knowledge bases, data assets, and API operations.

c. Each package supports a maximum of 3 discovery meetings per business scenario.

2) Technology landscape analysis: An overview of the current technological landscape for Large Language Models (LLMs), including the following:

a) Analyze the customer's current compute resources, network environment, and technical architecture.

b) Define the access policies and interface specifications for enterprise knowledge sources, databases, and applications.

c) Each package includes a technical integration assessment for up to 5 major systems.

3) The design of the agent technical pipeline includes the following:

a. Provide an agent-based technical solution, including the agent orchestration workflow, node design, model selection, and evaluation methods and standards.

b) Designing agent engineering pipelines, developing multi-node or multi-agent workflow solutions for complex scenarios, and providing technical consultation and guidance on agent development.

c) Each package supports one link design plan with a maximum of 5 nodes.

4) Scenario-oriented prompt optimization service, including the following:
a) Rewriting prompts for compatibility with the Qwen model.
b) Evaluating prompt performance and performing post-editing (PE) tuning.
c) Each package supports up to 3 rounds of prompt optimization. Each round is limited to the evaluation of 5 models and 1,000 evaluation data entries.

a. Rewrite Qwen model prompts to ensure compatibility.

b. Evaluation of prompt performance and optimization of PE.

c) Each package supports up to 3 rounds of prompt optimization, and each round is limited to 5 models and 1,000 evaluation data entries.

5) Project management services during project execution.

5.2 Agent building technical service items

Provides the following agent building technical services:

1) Research and analysis of requirements for Large Language Model (LLM) scenarios, including the following:

a) Interview the customer and confirm their business pain points and Large Language Model (LLM) application requirements.

b) Assess available resources, such as knowledge bases, data assets, and API operations.

c. Each package includes up to three research meetings per business scenario.

2) Current technology landscape analysis: Analyzes the current technology landscape of Large Language Models (LLMs), including the following:

a) Analyze the customer's current compute resources, network environment, and technical architecture.

b) Define the access capabilities and interface specifications for enterprise knowledge sources, databases, and application systems.

c) Each package includes technical integration assessments for up to five major systems.

3) Design and implement the technical pipeline for the agent, which includes the following:

a) Provide a technical solution for agents, detailing the orchestration workflow, node design, model selection, and evaluation methods and standards.

b) Design the MCP tools or Skills required for the agent engineering process and define their input and output standards.

c) Each package includes up to 1 technical link, 5 agent nodes, and 3 MCP tool/skill calls.

4) Contextual engineering optimization, including the following:
a) Designing the prompts for the nodes in the orchestration plan and performing multiple rounds of optimization.
b) Building and implementing the end-to-end engineering workflow and integrating it with your existing tools or Skills. This is done using Alibaba Cloud MaaS products in combination with agent orchestration solutions and contextual engineering.
c) Each package supports up to 5 rounds of prompt optimization and validation for up to 5 models.

a) Design the prompts for the nodes in the orchestration plan and iteratively optimize them.

b) Build and implement the end-to-end engineering pipeline and integrate it with the customer's existing tools or Skills using Alibaba Cloud MaaS products, agent orchestration solutions, and context engineering.

c) Each package supports up to 5 rounds of prompt optimization and validation for up to 5 models.

5) Agent evaluation, validation, and performance optimization, including the following:
a) Providing agent performance validation and high-availability (HA) deployment solutions. We also support you in completing the deployment of a production-level agent.
b) Providing bad case collection and optimization services and performing performance tuning based on the agent architecture, which excludes model fine-tuning. The final delivered performance is subject to the standards mutually agreed upon by you and Alibaba Cloud.
c) Each package supports a maximum of 2,000 evaluation data items, 5 rounds of bad case collection and optimization, and evaluation and optimization validation for up to 5 models.

a) Provides solutions for agent performance validation and high-availability (HA) to support production-level agent deployments.

b) Provides optimization services by collecting failure cases and performing performance tuning based on an agent architecture (excluding model fine-tuning). The final performance is subject to standards mutually agreed upon by Alibaba Cloud and the customer.

c) Each package is limited to 2,000 evaluation data items, 5 rounds of bad case collection and optimization, and 5 models for evaluation and optimization validation.

6) Project management services during project execution.

5.3 Model fine-tuning and training service items

Provides the following model fine-tuning and training services:

1) Requirements and scenario research: We provide research and analysis services for LLM fine-tuning or training scenarios, including the following:
a) Interviewing stakeholders to identify your business pain points and LLM fine-tuning or training needs.
b) Collecting and researching your available resources, such as existing training data, annotation resources, and computing power budgets.
c) Each package supports up to 3 research meetings for a single fine-tuning or training scenario.

2) Data and computing power assessment: We analyze your data quality and assess your computing power resources. This service includes the following:
a) We analyze the quality and format of the dataset you provide and offer recommendations for data formatting and sample quality checks. You are required to provide the dataset.
b) We assess the type and amount of resources required for your fine-tuning or training.

c) Pre-processing collected customer data as needed, including deduplication, filtering, and augmentation.

d) Each package includes a technical integration assessment for up to 5 main data sources, with a data volume of up to 50,000 samples.

3) Plan design and training task implementation, including the following:
a) Delivering a technical plan for fine-tuning or training. This includes base model selection, fine-tuning or training strategies such as LoRA, SFT, or Reinforcement Learning from Human Feedback (RLHF), hyperparameter design, and evaluation standards.
b) Configuring the training script and executing the training task based on the computing resources you provide.
c) Each package supports up to 1 training task, 1 model version, and 1 algorithm. The number of model training iterations cannot exceed 3.

4) Model evaluation, validation, and performance optimization, including the following:
a) Design core metrics that align with customer evaluation goals, such as accuracy, relevance, compliance, and response completeness.

b) Evaluate and deploy the model, and generate a model evaluation report using the evaluation dataset that you provide.

c) Provide

This offering includes failure case collection and optimization services. Performance is tuned based on the fine-tuning or training architecture. The final performance is measured against standards mutually confirmed by you and Alibaba Cloud.

d) Each package supports up to 50,000 evaluation data records, up to 3 evaluation iterations, and validation of evaluation and optimization for up to 3 model versions.

5) Project management services during project execution.

6.1 LLM consulting service

Application deadline: You must submit your application at least 15 calendar days before the service start date.

6.2 Agent building service

Application deadline: You must submit your application at least 15 calendar days before the service start date.

6.3 Model fine-tuning and training service

Application deadline: You must submit your application at least 15 calendar days before the service start date.