Exploring Gen AI — A Guide to Model Parameters for Better Content Creation (Part 2)

Source From Google

We are already aware of the Firebase Genkit & How to create a basic Firebase Genkit App & How to model the parameters to control the length of response we are getting from our model.

Build your First AI application with Firebase Genkit 🚀(Part 1)
https://shorturl.at/891Js

Build your First AI application with Firebase Genkit 🚀(Part 2)
https://shorturl.at/zPMCf

Exploring Gen AI — A Guide to Model Parameters for Better Content Creation (Part 1)
https://shorturl.at/BHA9u

Let’s check how to model parameters that control creativity.

Parameters that controls Creativity

1. temperature
   - This parameter is critical component in AI generated text that determines how random & creative the output is by controlling the probability distribution of model’s token predictions
   - In general, models predict the next word or token in a sequence by assigning probabilities to each possible next token
   - A low temperature makes the model focus on the most likely tokens, leading to deterministic & predictable output
   - A high temperature makes the model more likely to choose less probable tokens, resulting in creative & diverse outputs
   - Example Prompt: Describe a calm evening by the lake.
   The model predicts the following token probabilities:
   "peaceful": 0.5
   "tranquil": 0.3
   "quiet": 0.1
   "stormy": 0.05
   "noisy": 0.05
   - With Low Temperature (0.2) Result, The calm evening by the lake was peaceful and serene.
   - With High Temperature Result, The calm evening by the lake was tranquil, with whispers of an enigmatic breeze
2. topP
   - This parameter used in AI models to control randomness and creativity in output
   - It works dynamically selecting tokens based on their cumulative probabilities, ensuring a balance between creativity
   - While generating the next word or token LLM’s assign probabilities to each possible next token
   - Then, Top-P limits the token selection to only those whose cumulative probabilities add up to a specific threshold like 0.8 or 0.9
   - This ensures that model focuses on subset of likely tokens while ignoring highly improbable ones creating contextually appropriate responses
   - Example Prompt: Describe a magical forest.
   The model predicts the probabilities of the next tokens:
   "enchanted": 0.4
   "mystical": 0.3
   "dark": 0.2
   "haunted": 0.05
   "ancient": 0.05
   Then, Cumulative Probability:
   "enchanted": 0.4 (40%)
   "mystical": 0.4 + 0.3 = 0.7 (70%)
   "dark": 0.7 + 0.2 = 0.9 (90%)
   "haunted": 0.9 + 0.05 = 0.95 (95%)
   "ancient": 0.95 + 0.05 = 1.0 (100%)
   When top-P is applied:
   Top-P = 0.5: Only tokens contributing to the first 50% of the probability are considered ("enchanted", "mystical").
   Top-P = 0.9: Tokens contributing to the first 90% are considered ("enchanted", "mystical", "dark").
3. topK
   - Now we know language models assign probabilities to each possible next token in sequence
   - Top-K limits the selection to the K most probable tokens from the model’s predictions
   - From this Top-K tokens one is selected either deterministically or through sampling
   - Exactly K tokens are selected regardless of their cumulative probabilities
   - Example Prompt: Describe a magical forest. The model predicts
   "bright": 0.5
   "golden": 0.3
   "red": 0.1
   "cloudy": 0.05
   "dark": 0.05
   For top-K= 1, The sunrise was bright and warm.
   For top-K=3, The sunrise was bright and glowing

To adjust the parameters, update the code as given below

// import the Genkit and Google AI plugin libraries
import { gemini15Flash, googleAI } from '@genkit-ai/googleai';
import { genkit } from 'genkit';
import dotenv from 'dotenv';

// Load the API key from environment variables
dotenv.config();

// configure a Genkit instance
const ai = genkit({
  plugins: [googleAI({ apiKey: process.env.GOOGLE_GENAI_API_KEY })], // Pass the API key
  model: gemini15Flash,  // Set the model
});


(async () => {
  // make a generation request
  const { text } = await ai.generate({
    prompt: 'Write a description of Generative AI',
    config: {
    temperature: 1.5,      // 0-2
   topP: 0.9,           // Consider tokens with cumulative probability of 90%
    topK: 10            // Consider the top 10 most likely tokens
    }
  });
  console.log(text);
})();

Feel free to change the parameters and see the amazing results for getting the most creative and diverse output by keeping the balance in response.

🎉🎆🎇🎉🎆🎇🎉🎆🎇🎉🎆🎇🎉🎆🎇🎉🎆🎇🎉🎆🎇

What Next ?
Learn how to use special Genkit functions, called flows.

Stay Tuned and Keep Watch for Next Part, Happy Learning.
Feel Free to Clap 👏 and Keep me Motivated 👏 to come up with Next Parts.