Note_Tech

All technological notes.

---

Project maintained by simonangel-fong Hosted on GitHub Pages — Theme by mattgraham

AI Project Lifecycle

Back

• AI Project Lifecycle
  • Lifecycle
    • Scope of Program
    • Model Selection
    • Adapt and Align Foundation Model

---

Lifecycle

Scope of Program

• Identify the Use Case
• Define key objectives & outcomes

---

Model Selection

• FM model selection criteria

1. Modality : Text

   Image (Vision)

   Embedding - Command, Anthropic……

2. Size : Number of model parameters > 50B
3. Inference Speed or Latency : Response time for completion – Few Seconds
4. Context Window : 77- 200K token size – Claude has max context window
5. Pricing : FM pricing – Claude most expensive and Titan least
6. Training Dataset – Internet, Code, Human Feedback – Diverse dataset
7. Propriety or Open Source – Prefer Open Source
8. Fine-tunable – Should be fine-tunable
9. Additional Features – Multi-Lingual support – Jurassic, Titan
10. Quality of Response - Accuracy, Toxicity and Robustness

---

• Evaluate and select the Foundation Model

1. Automatic
   • Task Type: Text Generation, Text Summarization, Q&A, Text Classsification
   • DataSet: Built in – Gigaword, XSUM or Bring your own dataset
   • Metrics: Accuracy, Toxicity and Robustness
2. Human: Bring your own work team
   • Evaluates up to 2 models using a work team of your choice to provide feedback.
   • Provides results based on the parameters that are specified while creating the evaluation.
3. Human: AWS Managed work team

---

Adapt and Align Foundation Model

• Prompt Engineering
• RAG

• Fine Tuning

• Deployment & Integration
  • Deploy the Model
  • Integrate with existing
  • application landscape

---