Page 127 - AI for Good-Innovate for Impact Final Report 2024

AI for Good-Innovate for Impact

• UC36-REQ-08: Pipeline - Source of data (domain data) -> Collection mechanism -> Data
preprocessing -> Model training-> Supervised Learning (expert supervision) -> Fine
Tuning-> Distribution mechanisms.
• UC36-REQ-09: Model - A variety of algorithmic models, including the Transformer-based
Bert and the current mainstream large language models, ensuring the accuracy of the 27 - GRG
output results through the strategic integration of multiple models.
• UC36-REQ-10: Model training and finetuning - Instructions Tuning, LoRA Tuning.
• UC36-REQ-11: Case studies - This case study has been launched and is operational in
both new line projects and retrofit projects of the Guangzhou Metro and Shenzhen Metro.
• UC36-REQ-12: Testbeds/Experimentation/pilots/simulations/validations/tests - Model
unit testing, integrated testing of the overall solution.
• UC36-REQ-13: Metrics, KPIs, measurements - Accuracy rate of question and answer
within the scope of domain knowledge.
• UC36-REQ-14: Use case scenarios and Requirements - Passenger ticket purchasing
scenarios in rail transit travel, scenarios where passengers consult with station staff, etc.
• UC36-REQ-15: Role of Trainings, standards - Enhance the generalization ability and
application effectiveness of general large models in specific domains and downstream
tasks.
• UC36-REQ-16: Role of open source - Accelerate model iteration, enhance the
foundational capabilities of the model.

27�4� Sequence diagram

111