Tutorial: What is Generative AI and How Can It Be Applied?
Generative AI, a cutting-edge technology, refers to deep-learning models capable of producing high-quality content ranging from text to images. One significant type of generative AI is the Large Language Model (LLM), known for its general-purpose understanding and generation capabilities.
In this tutorial, we will explore how to leverage prompt engineering with LLMs to obtain accurate, relevant, and contextually rich responses, specifically related to travel information on various countries.
Step 1: Getting Started
Open Watsonx.ai Prompt Lab and select the Freeform mode option. You will see the prompt editor in the center, alongside model parameters on the right side for optimizing responses. A token summary appears on the bottom-left corner, indicating the number of tokens used during execution.
Note that the models are foundation models hosted on the IBM Cloud, allowing for model inference and interaction.
Step 2: Initial Prompt
Begin by entering a simple prompt. For instance:
Model: flan-t5-xxl-11b
Prompt text: I am thinking of traveling to Thailand.
The output may not yield useful information, akin to asking a vague question with broad answers. Now, we will refine the prompt to gather more specific data.
Step 3: Refining the Prompt
To get better results, make the prompt more direct:
Prompt text: I am thinking of traveling to Thailand. Tell me about Thailand.
This version produces a more relevant response but may end abruptly due to Max tokens parameter constraints. Increase the maximum tokens to improve this.
Step 4: Adjusting Model Parameters
Increase the Max tokens to 200:
Model: flan-t5-xxl-11b
Prompt text: I am thinking of traveling to Thailand. Tell me about Thailand.
This will allow for a complete response. If the model consistently returns the same answer, switch the decoding mode to Sampling instead of Greedy Decoding to generate varied outputs.
Step 5: Targeted Information
To tailor responses to user interests, refine the prompt further:
Prompt text: I am thinking of traveling to Thailand. I like water sports and food. Tell me about Thailand.
If information remains limited, consider exploring alternative models.
Step 6: Exploring Other Models
Watsonx.ai provides model cards with comprehensive details on various models. Access this by clicking on the dropdown near the model name:
- Provider and source information
- Best-suited tasks for the model
- How to tune the model
- Research white papers
- Bias, risks, and limitations
Consider selecting the llama-2-70b-chat model, which is optimized for dialogue use cases.
Step 7: Testing a New Model
With the new model selected, rerun the previous prompt:
Prompt text: I am thinking of traveling to Thailand. I like water sports and food. Tell me about Thailand.
Ensure you monitor the output length and avoid hitting the Max tokens limit again.
Step 8: Adding Limits to Responses
Add constraints directly to the prompt for more focused results:
Prompt text: I am thinking of traveling to Thailand. I like water sports and food. Give me 5 sentences about Thailand.
These modifications will lead to tailored, informative results while staying within your token budget.
Conclusion and Next Steps
Prompt tuning is a practical alternative to training new models for specific needs. This tutorial illustrated the iterative nature of prompt engineering, emphasizing the importance of context and targeted queries.
As a final note, all model interactions occur through the IBM Cloud. For technical details about the API calls, review the output by selecting View code in the prompt lab.
For further learning, consider experimenting with different prompts and parameters to discover nuanced, user-focused generative responses!
Lasă un comentariu
Toate comentariile sunt moderate înainte de a fi publicate.
Acest site este protejat de hCaptcha și hCaptcha. Se aplică Politica de confidențialitate și Condițiile de furnizare a serviciului.