Page 761 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

• Engaging with refugee communities and local stakeholders to ensure the project meets
their needs and includes their feedback.

Working with technology providers like OpenAI or Meta to leverage their expertise and
resources for model development and fine-tuning. 4.9: Accessibility

3 Use Case Requirements

REQ-01: It is critical that sufficient processing power and compute time are allocated to support
the fine-tuning of OpenAI’s Whisper Large model. Given the training data exceeds 400 hours of
audio, high-performance GPUs such as the RTX 4080 are required, with an estimated compute
time of 200–400 hours depending on the number of training epochs.

REQ-02: It is expected that dedicated technology infrastructure is established to host and
serve the fine-tuned language models. This includes setting up a secure and scalable server
environment with appropriate access configurations and safeguards to ensure reliable system
availability.

REQ-03: It is expected that model performance is continuously improved through the use of
advanced machine learning techniques. This includes experimenting with larger transformer
architectures and multilingual pre-trained models to enhance language understanding and
accuracy.
REQ-04: It is critical that the project is supported by a skilled interdisciplinary team. Required
roles include NLP specialists, assistant researchers, software developers (front-end and back-
end), User Interface/User Experience(UI/UX) designers, and software Quality Assurance(QA)
testers to ensure robust development, integration, and user experience.

4 Sequence Diagram

725