Guiding unstructured responses

Working with Llama 3

Imtihan Ahmed

Machine Learning Engineer

Controlling model output

✅ Parameters
✅ Roles

How to refine model

Controlling model output

✅ Parameters
✅ Roles

How to refine model - refine prompts

Controlling model output

✅ Parameters
✅ Roles

How to refine model - zero-shot/few-shot prompting

Controlling model output

✅ Parameters
✅ Roles

How to refine model - use stop words

Refining prompts

Example: summarization

text_choice1 = "Summarize key trends in the aviation industry from the last year, 
                focusing on fuel efficiency innovations."


text_choice2 = "Tell me about the aviation industry."


output = llm(text_choice1) # More specific prompt is more effective

print(output['choices'][0]['text'])

The aviation industry has made significant strides in fuel efficiency innovations 
over the last year, driven by the need to reduce greenhouse gas emissions and 
operating costs. Sustainable Aviation Fuels (SAFs) have emerged as a key trend...

Components of effective prompting

Components of effective prompting - Precision

Components of effective prompting

Components of effective prompting - ambiguity

Components of effective prompting

Components of effective prompting - keywords

Components of effective prompting

Components of effective prompting - action words

Components of effective prompting

Components of effective prompting - examples

Zero-shot prompting

Zero-shot prompting: a single instruction

text = "Summarize recent mergers in the airline industry."


output = llm(text) 

print(output['choices'][0]['text'])

Recent mergers in the airline industry include Alaska Air Group's acquisition of 
Hawaiian Airlines in 2024, with both airlines continuing to operate as separate 
brands. In 2022, Delta Air Lines purchased 20% of LATAM Airlines Group...

Refining zero-shot prompts

Need to distinguish task, expected output, additional context
Zero-shot prompting with labels

text = """
       INSTRUCTION: Write concisely and in 2-3 sentences that cover only key points.
       QUESTION: Summarize recent mergers in the airline industry.
       ANSWER:
       """

Few-shot prompting

Few-shot prompting: use of multiple examples

Prompt

Few-shot prompting

Few-shot prompting: use of multiple examples

Prompt with 'aircraft model' highlighted

Few-shot prompting

Few-shot prompting: use of multiple examples

Prompt with 'passenger capacity' highlighted

Few-shot prompting

Few-shot prompting: use of multiple examples

Prompt with 'fuel consumption' highlighted

Few-shot prompting

output = llm(f"Continue the entries: {text}") 

print(output['choices'][0]['text'])

Aircraft Model: Airbus A350-900
Passenger Capacity: 350
Fuel Consumption: 2.39 liters per seat per 100 km

Using stop words

Need concise insights
Use stop words to end the response at a specific point
Example: question-answering application

text = "Which airlines operate direct flights from London to Singapore?"


output = llm(text, stop=["Q:"]) # Stop responses at "Q:"

print(output['choices'][0]['text'])

You can fly direct from London to Singapore with Singapore Airlines and 
British Airways.

Let's practice!

Working with Llama 3