When you follow technical blogs and news tickers you may have read about BLOOM in the context of AI and text generation. I became curious and within this article I want to share my findings with you.
If you are not interested in background information and just want to start with your experiments then you can jump to the installation instructions.
So what is BLOOM you ask? You may already have heard of transformer based large language models like Gopher, GPT-3, LaMDA and others.
Trained on massive amounts of texts such systems can produce/generate new text based on a short input string often described as "prompt".
For a lot of usecases the resulting texts turn out to appear somehow smart which is why such systems are often used and evaluated in the context of Artificial Intelligence (AI).
Use cases include building chatbots, generate articles and blog entries based on a list of keywords or support developers while coding. For example GitHub CoPilot became a commercial product already which proves you can make money with such a technology.
It would be an omission not the mention that some systems attracted attention for creating biased content - which is of course due to the material those systems have been trained with.
As a user such problems are difficult to predict as more or less everything what happens behind the user interface appears as a black box. You may be able to rate the generated results but if they appear inappropriate there is not really a way to debug the scenario.
Even worse: results may appear well argued and (syntactically) correct so therefore you may feel tempted to rate the system response as valid. But how can you be sure? The answer is: you can't.
So when working with such systems you will find yourself looking for prompts that create meaningful results. On discussion forums you can already find lists of "prompts" that have been proven to work for a particular system.
Correct, I tried to draw a picture first to help you understand what may be different with BLOOM appearing on stage.
Some articles draw the conclusion that BLOOM might be "the Most Important AI Model of the Decade". Others call it a "Gamechanger".
Among the list of reasons what stands out is that the creators of BLOOM follow an "open" approach from the very beginning. This is not a product from a single company but the result of work from an international team.
An Ethical Charter has been defined as a guideline which helps to make decisions both transparent and comprehensible.
The makers of BLOOM describe the result as follows:
"BLOOM is an autoregressive Large Language Model (LLM), trained to continue text from a prompt on vast amounts of text data using industrial-scale computational resources. As such, it is able to output coherent text in 46 languages and 13 programming languages that is hardly distinguishable from text written by humans. BLOOM can also be instructed to perform text tasks it hasn't been explicitly trained for, by casting them as text generation tasks." Source
And more:
"any individual or institution who agrees to the terms of the model’s Responsible AI License (developed during the BigScience project itself) can use and build upon the model on a machine or on a cloud provider" Source
When the creators of BLOOM promise that a developer "can use and build upon the model on a local machine" the reader might belive that there is a way to simply download and install the whole environment like you would do with tools as Visual Studio Code. But it turns out that things don't work like that. The whole system being open does not mean that you have the processing power available to use BLOOM offline in any reasonable way. Some people tried it though and you may be interested to follow this thread in a discussion board.
You want to retrain the model? Before you start have a look at this and make sure you have the appropriate hardware available 😊
If you - like my - are just interested in how the technology feels and behaves, if you want an easy approach to get things up and running, if you want to play around with some prompts then there is another way: a public API allows quick tests, prototyping, and lower-scale use.
There are quite some examples around how to make use of this API. Most of what you will find on the web is Python based code. If you feel more familiar with NodeJS and want to have something up and running in under an hour then this project is for you.
For this example to work you need:
package.json" and "index.js" from this projectindex.js and set the constant apiToken with the value you created abovenpm installnode index.jsIn under a minute you should see a result in your terminal.
There is a lot of mathematics going on in the background but from a user perspective what is happening is that you define a so called "prompt" which is the initial string parameter for your request. Bloom will then try to create a completion string for your initial prompt. In a lot of cases the results you get make sense but of course you will also run into situations where the responses are completely useless.
The big challenge when working/programming with this approach is to work out prompts that create meaningful results.
In addition to finding a good prompt you also have to decide about values for a long list of available parameters there are a lot. Modifying parameter settings will influence the results you get.
When working with BLOOM currently both prompt length and result length are limited in terms of size. With GPT-3 I was able to create long texts. This will not work with BLOOM. Hopefully they will change this restriction in the near future.
Furthermore a resulting completion string may stop right in the middle of a sentence. Not only because of this it seems necessary to filter and postprocess all results in case you plan to use them in a real world application.
The API is not meant to be fast: it may take a minute or so to get a result. So do not plan to built a real time based on this approach.
Finally here are some examples for prompts and completions.
When you run these examples yourself you may get different results because of different parameter settings (which I did not document here). Furthermore running the same request twice may produce different results.
Hint: please also note that I manually reformatted some results for better readability.
Do you know these riddles where you are asked to complete a list of given number? Let us see if BLOOM can help us with that, we start with some prime numbers:
Prompt: "2,3,5,7,11,13"
Completion: "2,3,5,7,11,13,17,19,23,29,31,37,41,43,47,53"
Well done. Continue with some Fibonacci numbers:
Prompt: "1,1,2,3,5,8,13"
Completion: "1,1,2,3,5,8,13,21,34,55,89,144,233,377,610,987,15"
The last number should be "1597" but BLOOM threw away the last two digits ...
How about a sequence of even numbers?
Prompt: "2,4,6,8,10,12"
Completion: "2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32"
Let us introduce a pattern of a switching sign:
Prompt: "2,-4,6,-8,10,-12"
Completion: "2,-4,6,-8,10,-12,14,-16,18,-20,22,-24,26,-28,30,-32"
We can also do text based math:
Prompt: "Sum of integers from 1 to 10 is"
Completion: "Sum of integers from 1 to 10 is 55"
Wow, that is correct.
Prompt: "What is the square root of nine"
Completion: "What is the square root of nine? Three."
Now replace "nine" with "9" and let us see what happens:
Prompt: "What is the square root of 9"
Completion: "What is the square root of 9? 3."
That works too. We try the same with "-9"
Prompt: "What is the square root of -9"
Completion: "What is the square root of -9? I don't know."
Sometimes you have some samples and want to complete a set with more items.
Greek alphabet:
Prompt: "alpha beta gamma delta"
Completion: "alpha beta gamma delta epsilon zeta eta theta iota kappa lambda mu nu xi omicron pi rho sigma tau"
Let us leave away the "delta" in the prompt and try again.
Prompt: "alpha beta gamma"
Completion: "alpha beta gamma", "alpha beta gamma", "alpha beta gamma", "alpha beta gamma","
Obviously three examples where not enough for BLOOM to guess the pattern. But it helps if we add a hint to the prompt:
Prompt: "Complete list: alpha beta gamma"
Completion: "Complete list: alpha beta gamma delta epsilon zeta eta theta iota kappa lambda mu nu xi omicron pi rho sigma"
So after adding the term "Complete list" to the prompt we received a valid result again. Remember as said above A lot of working with such APIs is about finding an appropriate task description.
Prompt: "Complete list: red,green,blue"
Completion: "Complete list: red,green,blue,yellow,orange,purple,pink,brown,grey,black,white,"
The model was trained for a lot of languages. German is not on the list. But we will give it a try though:
Prompt: "Complete list: rot,gelb,grün"
Completion: "Complete list: rot,gelb,grün,blau,lila,orange,magenta,violett,weiss"
This is indeed a valid result for German.
Will list completion work with actors of a famous TV series?
Prompt: "Complete list: Picard,Worf,Data"
Completion: "Complete list: Picard,Worf,Data,Geordi,La Forge,Ro Laren,Tasha Yar,Wesley Crus"
We see the system was unable to add "Wesley Crusher" to the list. I understand that. This may be even a feature.
We switch the focus to generate some text. As mentioned above currently the response size is limited which may change in the future. But we give it a try:
Prompt: "Short story: Once upon a time"
Completion: "Short story: Once upon a time there was a girl who was born in a small town in the south of France. She was born"
I'd like to know how the story continues - maybe next year BLOOM will create longer responses?
Will we soon be able to autocreate our articles blog posts in the future?
Prompt: "Article with keywords: java,POST,request"
Completion: "Article with keywords: java,POST,request POST request to a URL in Java using Apache HttpClient. You need to set the content type"
Another try:
Prompt: "Blog post: How to print 'hello world' in NodeJS"
Completion: "Blog post: How to print 'hello world' in NodeJS - https://medium.com/@jay.k.sharma/how-to-print-hello"
You don't need to try: the link does not work for me (404).
Prompt: "Article: How to print 'hello world' in Python"
Completion: "Article: How to print 'hello world' in Python - https://stackoverflow.com/questions/381278/how-to-print-hello-world-in-python"
This time the link exists, but the article is unrelated to the prompt.
If only short texts are allowed how about generating tweets for Twitter?
Prompt: "Hashtags: elon, musk, funny. Tweet:"
Completion: "Hashtags: elon, musk, funny. Tweet: Why Not: SpaceX #SN106 Elon's SpaceX deployed its boilerplate Falcon 9 Rocket today from NASA's historic Launch complex 39A at Kennedy Space Center with a new HttpClient"
I agree this is somehow funny.
Prompt: "Hashtag list: Alexa, joke, skill. Tweet:"
Completion: "Hashtag list: Alexa, joke, skill. Tweet: Alexa, stop Official: That's not available at the moment"
I am curious. What does BLOOM know/think about myself?
Prompt: "Who is Frank Börncke"
Completion: "Who is Frank Börncke and why is he so important? He is the head of the German Association of the Automotive Industry"
I can assure you this is not true! Let us rephrase the prompt:
Prompt: "The person Frank Börncke is"
Completion: "The person Frank Börncke is currently a professor at the University of Oslo, Norway. He is also a member of the Norwegian Academy"
I did not know that yet.
Prompt: "Frank Börncke is located in"
Completion: "Frank Börncke is located in Tübingen, Germany. He holds a Master in Economics and a PhD in Business Administration."
For the record: I deny that this is true. I am located in the Munich area. I continue with a question about my family:
Prompt: "Frank Börncke's wife's name is"
Completion: "Frank Börncke's wife's name is Susanne Börncke. She is a teacher. She teaches at a school in the village"
Please believe me: I do not know this woman.
Recommend a Smart Speaker:
Prompt: "Recommendation: The best smart speaker model is"
Completion: "Recommendation: The best smart speaker model is the one that fits your needs. If you want a speaker that can do everything, then the Google"
Though syntactically this is a valid answer I do not agree with the conclusion.
Prompt: "Recommendation: The best car is"
Completion: "Recommendation: The best car is the one that you can afford."
I like this answer.
Let us try to wrap a translation task into a prompt:
Prompt: "Translation from english into french. Input: 'I love you' Output:"
Completion: "*Translation from english into french. Input: 'I love you' Output: 'Je t'aime'
Another one:
Prompt: "Translation from english into french. Input: 'Peter loves Mary' Output:"
Completion: "Translation from english into french. Input: 'Peter loves Mary' Output: 'Peter aime Marie'"
Create lists by description:
Prompt: "List: famous Depeche Mode Songs:"
Completion: "*List: famous Depeche Mode Songs:
We choose another Artist:
Prompt: "List: famous Madonna Songs:"
Completion: "*List: famous Madonna Songs:
List some male actors:
Prompt: "List: famous actors (male):"
Completion: "*List: famous actors (male):
Why only John all the time? We don't know. We continue with fruits:
Prompt: "List: fruits:"
Completion: "*List: fruits:
Prompt: "List: prime numbers:"
Completion: "List: prime numbers: 5, 7, 11, 13, 17, 19, 23, 29, 31, 37,"
Looks incomplete for me at the beginning. But at least not any false entries in the list.
Can we ask for shell commands by describing a task?
Prompt: "Instruction bash ls command: show files in current directory including creation date"
Completion: "Instruction bash ls command: show files in current directory including creation date: ls -lrt"
Another one:
Prompt: "Instruction bash command: show biggest file in current directory:"
Completion: "Instruction bash command: show biggest file in current directory: ls -l | awk '{print $5}' | sort -nr | head -1"
Not exactly what I asked for but still an impressive result.
Will this work for SQL statements too?
Prompt: "Instruction : sql statement how to find the maximum price in table items:"
Completion: "Instruction : sql statement how to find the maximum price in table items: SELECT MAX(price) FROM items;"
Here is a very generic answer for a very generic question:
Prompt: "Recipe: How to make a cocktail:"
Completion: "*Recipe: How to make a cocktail:
As the foregoing examples have shown there are a numbers of use cases where BLOOM can be helpful. But there are also results that are completely unusable. Therefore currently it is necessary to validate all results that you get for plausibility before delivering them to a user.
The challenge when working with tools as BLOOM is to find prompts and parameter settings that deliver results that match your request. We also saw that minimal changes in the prompt can lead to different results.
A big limitation for real world applications compared to GPT-3 is the length/size limitation, especially for results. I hope this will change in some time.
My expectation is that without access to better tools programming with text completion AI frameworks like BLOOM or GPT-3 will feel like trial and error while searching for appropriate prompts.
And I'd like to foretell that in some time we will see more and more requests on stackoverflow.com discussing prompt candidates that lead to expected results. Wee will see.