Course 06

Align Your Model

This course tackles the critical challenge of making AI systems helpful, harmless, and honest. Learn to treat alignment as a technical engineering discipline, including RLHF, preference learning, bias evaluation, and safety techniques for African contexts.

Key Topics

  • AI Alignment: Defining human values and intentions
  • RLHF (Reinforcement Learning from Human Feedback)
  • Mitigating toxicity and bias
  • Safety evaluations and Red Teaming
  • Cultural sensitivity in African contexts

Learning Outcomes

  • Explore techniques to ensure model outputs align with human intent
  • Understand technical methods for suppressing harmful outputs
  • Integrate ethical considerations into the engineering lifecycle
  • Implement alignment and safety techniques
  • Address algorithmic bias technically

Practical Projects

  • Train a Safety Guardrail model for African language chatbots
  • Evaluate model bias using African stereotype datasets
  • Implement DPO (Direct Preference Optimization)

Recommended Datasets

3 datasets are aligned with this course for hands-on exercises and projects.

AfriHate

Hate Speech Detection for 15 African Languages

hate-speech-detection text-classification ~45,000 examples
Algerian Arabic Amharic Hausa Igbo +8 more

AfriStereo

African Stereotype Evaluation Dataset

bias-evaluation stereotypes ~5,000 pairs
English (African context) French

OASST1

OpenAssistant Conversations for RLHF

rlhf dialogue 161k messages
35 languages (Multilingual)

About This Course

As AI models become more powerful, ensuring they align with human values becomes critical. This course covers techniques for making models helpful, harmless, and honest.

Key Topics

  • Alignment Fundamentals: Why alignment matters
  • Preference Learning: Learning from human preferences
  • RLHF: Reinforcement Learning from Human Feedback
  • Safety Evaluation: Red teaming and safety benchmarks
  • Responsible AI: Ethical considerations in deployment

Prerequisites

  • Completion of Courses 01-05 or equivalent knowledge
  • Understanding of reinforcement learning basics
  • Familiarity with evaluation metrics