Code Execution with Google Gemini Flash

Introduction

Massive Language Fashions, the successors to the Transformers have largely labored throughout the house of Pure Language Processing and Pure Language Understanding. From their introduction, they’ve been changing the standard rule-based chatbots. LLMs have a greater potential to know textual content and might create pure conversations, so they’re changing the traditional chatbots. However since their introduction, the LLMs are doing greater than what they’re able to. Like changing Pure Language to SQL Queries, in a position to browse the web to fetch the newest data. And now they’ve the flexibility even to execute code. On this article, we are going to have a look at the newly launched function of Gemini, i.e. the Code Execution.

Studying Aims

• Study Code Execution with LLMs.
• Get launched to Gemini Flash 1.5.
• Learn to get the API Key for Gemini.
• Understanding how the LLMs fail in mathematical duties.
• Leveraging LLMs with Code Execution for exact and correct solutions.

This text was revealed as part of the Knowledge Science Blogathon.

Gemini – Google’s Massive Language Mode

Gemini fashions are a household of enormous language fashions launched by Google. It’s launched by Google to rival the favored closed-source massive language fashions just like the GPT4 from OpenAI and Claude from Anthropic. Gemini is a multimodal massive language mannequin that’s able to understanding textual content, pictures, audio, and even movies.

GPT4 was in a position to do the identical as Gemini does however one which it differentiates from Gemini is working the Code that it generates. And now just lately Google has up to date the Gemini mannequin making it run code. The code execution is feasible because of the operate calling capabilities of Gemini, the Code Execution is similar to it and the code it generates, will run and get the outcomes to generate the ultimate output to the person.

The code that Gemini generates will probably be run in an remoted sandboxed surroundings. Proper now, solely the numpy and the sympy libraries are current within the sandboxed surroundings and the generated code on no account can obtain and set up new Python libraries.

Getting Began with Code Execution

Earlier than we start coding, we have to get the free API key that Google offers to check the Gemini mannequin. The free API even helps the Code Execution. To get the free API, you’ll be able to click on on the hyperlink right here. Now, we are going to begin with putting in the library.

!pip set up -q -U google-generativeai

It is very important preserve the -U flag whereas putting in the google-generativeai library. It is because the Code Execution is a brand new function and would require the newest model of google-generativeai library to work. Now we are going to authenticate ourselves.

import google.generativeai as genai

GOOGLE_API_KEY = "YOUR API KEY"

genai.configure(api_key=GOOGLE_API_KEY)

Right here we import the google.generativeai library and name the .configure() technique. To this, we give the API Key that we have now obtained by signing to the Google AI Cloud. Now we will begin working with the Gemini Fashions.

mannequin = genai.GenerativeModel(model_name="gemini-1.5-flash")

response = mannequin.generate_content("How are you?")

print(response.textual content)

Rationalization

• Right here we begin by creating an occasion of the GenerativeModel Class.
• Whereas instantiating this object, we give the identify of the mannequin that we’re working with, which right here is the gemini-1.5-flash, which is the newest mannequin from Google.
• To check the mannequin, we name the .generate_content() technique after which, give the question to it, and retailer the generated textual content within the response variable.
• Lastly, we print the response. We will observe the response in pic above.

Not all the things will be answered accurately with the Massive Language Mannequin. To check this, allow us to strive asking the Gemini Flash mannequin a easy query to show the primary 5 letters of the phrase Mississippi.

response = mannequin.generate_content("Trim this phrase to first 5 letters, Mississippi")

print(response.textual content)

Right here, working the code and seeing the output above, we see that Google’s Gemini mannequin, the newest LLM development from the Google group has didn’t reply such a straightforward query. This isn’t solely with the Google Gemini fashions, however even the GPT4 from OpenAI and even Claude from Anthropic fail to reply it.

It is because they don’t have the flexibility to depend backward. That’s after producing the letter “i” the mannequin has no concept that it has outputted the second letter. It simply outputs a letter given the earlier letter, however has no concept in regards to the size of the earlier letters.

One other Instance

Allow us to check out one other query that the massive language mannequin fails to reply.

response = mannequin.generate_content("What's the sum of first 100 fibonaocci numbers?")

print(response.textual content)

Right here, we ask the Gemini Flash mannequin to present us the sum of the primary 100 Fibonacci sequence. Working the code and seeing the output pic, we will say that the mannequin has didn’t reply our query. As a substitute of returning the sum, it has given us the steps to get the sum of the primary 100 Fibonacci sequence. The mannequin failed as a result of massive language fashions are text-generation fashions. They haven’t any potential to carry out mathematical operations

So in each instances, the mannequin has failed. Now, what if the gemini mannequin has potential to execute Python code? The mannequin might attempt to write a code that would lead us to the reply we expect. Possibly for the primary query, the mannequin might carry out a string operation and run the code and for the second query, it might write a operate to calculate the sum.

Gemini – Code Execution

So now, allow us to attempt to ask the mannequin the identical two questions however this time, offering it entry to the Code Execution instrument.

model2 = genai.GenerativeModel(model_name="gemini-1.5-flash", 
instruments="code_execution")

response = model2.generate_content("Trim this phrase to first 5 letters, 
Mississippi. Use code execution instrument")

print(response.textual content)

Right here once more, we create an occasion of the category GenerativeModel and provides it the Gemini-1.5-flash mannequin identify, however together with it, we even give it the instruments that the mannequin can work with. And right here we offer it with the code_execution instrument. Now, we ask the identical query to the mannequin. This time, we even inform it to work with the code_execution instrument.

Working the code and seeing the output pic above, we will discover that, the Gemini Flash mannequin has written a Python code to do a string operation i.e. slicing right here, it has sliced the primary 5 letters of the phrase Mississippi and has lastly given us the answered that we needed. Now allow us to strive the identical with the second query, the place we ask the LLM to present us the sum of the primary 100 Fibonacci numbers.

response = model2.generate_content("What's the sum of first 100 fibanocci numbers?")

print(response.textual content)

Right here, working the coding and seeing the output, we see that the Gemini Flash has generated a operate to calculate the Fibonacci quantity. Then known as the operate by giving it 100 for n worth after which lastly printed the output. With the code_execution instrument, the gemini llm was in a position to accurately give us the reply to the query. This fashion it may possibly clear up mathematical issues by making a code out of it and working the code to get the reply.

Conclusion

The introduction of code execution in Google’s Gemini mannequin represents a big development within the capabilities of enormous language fashions. By integrating this function, Gemini can not solely perceive and generate textual content but additionally execute code to unravel complicated issues. This growth enhances its utility in quite a lot of purposes, from pure language processing to performing particular computational duties. The flexibility to run code permits Gemini to beat a few of the inherent limitations of language fashions, significantly in dealing with exact calculations and procedural duties. 

Key Takeaways

• Gemini can perceive and course of textual content, pictures, audio, and video, making it a real multimodal.
• Massive Language Fashions typically fail to reply mathematical questions with precision, as a result of they can not carry out calculations.
• Code Execution permits an LLM to run code in a sandboxed surroundings.
• Massive Language Fashions can run Python Code by performing a instrument name and giving the instrument the related Python code to execute.
• Google’s free API permits customers to entry the Gemini Flash API that may Execute Code.

Ceaselessly Requested Questions

The media proven on this article will not be owned by Analytics Vidhya and is used on the Creator’s discretion.