Evaluations

Hallucinations

How to avoid them

1. Choose Advanced Models

2. Write Clear Instructions

Tell the AI: “Answer only with verified information. If you’re not sure, say you don’t know.”

3. Use Step-by-Step Reasoning

• Guide the model: “Let’s solve this step by step.”

• More Explicit. For example:

  • “First, identify the key variables.
  • Second, check what information is missing.
  • Third, calculate each component separately. Finally, combine the results.”

• For more control, you can implement this logic in code as a state graph

4. Provide Examples

• few-shot prompting

5. Ground with Real Data

• RAG

6. Lower the Temperature

7. Implement Self-Checks

• ask it to verify: “Are you sure? Can you double-check that information?”
• Voting mechanism

8. Add External Verification

• Rule based external validation

9. Fine-Tune on Your Domain

• Fine-tune on verified

10. Use Human Feedback

• Reinforcement Learning from Human Feedback

Evaluations

References:

1. LLM Eval FAQ, Hamel Husain Shreya Shankar, 2025-06-29
2. Huggingface Evaluations