Technical Development Service Agreement
Exhibit 10.9
Technical Development Service Agreement
Project Name: Vertical Pre-training Program for Artificial Intelligence Education – LLM Model
Agreement Parties:
Client (Party A): Dragonsoft Group Co., Limited
Technical Service supplier (Party B): GGA Technology Inc.
Location of Signing: Hong Kong
Date of Signing: October 2, 2023
After negotiation, the two parties have reached and entered into this agreement concerning the technical development (commissioned/joint development) of the Large Language Model (LLM) vertical pre-training program for the education industry.
1. Subject Matter, Form, and Requirements of the Technology:
Party A entrusts Party B to conduct pre-training and necessary fine-tuning of a general-purpose open-source AI large language framework model, in line with the needs of the education industry. After completion, the model will be deployed and tested on the platform designated by Party A, and provided for Party A's use. The computational resources required for the research, development, and training will be provided by Party B.
2. Technical Indicators and Parameters:
Based on the open source large model Llama2, pre-training will be conducted, incorporating the characteristics of English and Chinese language teaching, to form the aforementioned LLM large language model for the vertical field of English and Chinese education. The model should be independently deployable and applicable to related industry applications such as teaching, dialogue training, vocabulary exercises, semantic evaluation, and language ability assessment.
Foundational Model:
13B model of Llama 2 (or 70B model if needed)
Training Data:
(1) Publicly available mixed data from the open source training corpus of Llama 2;
(2) English/Chinese teaching training data collected and organized by Party B;
(3) The proportion of training data in Chinese language should not be less than 50%.
Other Inclusions:
Network Architecture: Standard Transformer,
pre-normalization: RMSNorm,
context length: up to 4k,
grouped-query attention: GQA.
Training Optimization:
Iterative optimization of the model through rejection sampling and proximal policy optimization (PPO). In reinforcement learning from human feedback, cumulative and iterative reward modeling data and model improvement shall be carried out in parallel.
Basis for testing and acceptance of this project:
(1) The dialogue effect of the industry large model is evaluated using a manual evaluation method.
(2) The evaluation includes multiple dimensions such as semantic understanding, topic control, role assignment, multi-turn dialogue, data analysis, and feedback speed.
(3) The assessment is primarily conducted in English and Chinese.
The actual functional use of this LLM for the education should be capable of achieving and meeting the client’s requirements for auxiliary teaching in English instruction (including English proficiency assessment), as well as auxiliary teaching and proficiency assessment in Chinese instruction.
3. Development Plan:
(1) From October 2023 to the end of January 2024: complete the LLM training, inspecting, fine-tuning and deployment.
(2) The computing power and resources used for the total AI model training will be provided by Party B.
4. Development and Service Expenses:
Research and development expenses refer to the costs required to complete the research and development and all related services. The total development expenses and remuneration for this project: a total of two million US dollars,
If there are changes to the execution and functions, the total payment amount can be adjusted accordingly.
Payment can be made in USD or HKD.
The payment method (Chose the second method): (1) The fee is paid in full in one lump sum; (2) The 60% of it to be paid before October 20, 2023, and the 40% of it to be paid before March 31, 2024.
5. Ownership of Equipment and Instruments with Development Expenses:
(1) The software code developed in this project belong to Party A (if any).
(2) This is a software development and service project and does not have any purchase of equipment and instruments.
6. Timing and Method of Performance:
(1) This agreement begins to be performed according to the development plan from the date of signing.
(2) The performance and acceptance of this agreement: The LLM/LLM-Chat pre-trained vertical model is deployed online on the designated platform and operates normally.
7. Confidentiality of Technical Information and Data:
Both parties collaborate to keep the development information confidential. Without the consent of the other party, no information related to this project may be disclosed to a third party.
8. Technical Cooperation and Direction:
(1) Party B is responsible for providing the development environment, training computing power, etc.
(2) Party A is responsible for testing and inspecting after the completion of the AI model training, and the using in application platform, etc.
9. Assumption of Risk and Responsibility:
(1) If any delay is caused by Party A's failure to arrange time to do system testing and inspecting in a timely manner, the responsibility shall belong to Party A.
(2) If the project development fails or partially fails due to objective limitations of existing technology and technological force majeure, the risk and responsibility shall be borne equally by both parties.
10. Ownership and Sharing of Technical Develop Results:
This project’s use rights, transfer rights of the developed system, documents, and technical secrets belong to Party A.
11. Standards and Methods of Acceptance:
The technical results completed by the development and service should meet the technical indicators listed in Article 2 of this agreement according to industry standards, and be confirmed by Party A for acceptance.
12. Governing Law:
The execution of this agreement is governed and protected by the laws of the Hong Kong Special Administrative Region of the People's Republic of China.
13. Dispute Resolution:
Disputes arising during the performance of this contract can be resolved through reconciliation or mediation by the parties. If the parties are unwilling to reconcile or mediate, or if reconciliation or mediation fails, the following method will be used to resolve: Litigation in the courts of the Hong Kong Special Administrative Region.
14. This agreement is in two copies, each party holds one copy, and they have equal legal force.
Signature:
| Party A: | Party B: | |
| Authorized Signature | Authorized Signature |
