license: wtfpl
language:
- en
tags:
- companion
- assistant
- question answering
- LittleMKIA
- MKIA
- chatbot
- useful
- personal model
- offline model
- friend
- privacy
MKIA The SpokesCat for this model
MKIA
stands for
Miss Know It All
and she is a Catbot not a chatbot.
I am releasing this model under the terms that people can do whatever they want to with it. All I ask is that they call any copies or derivatives of it by another name, otherwise, they can do whatever they want to with it.
Description
She is intended to be a companion and assistant. She is a fine-tuned version of Google's 780 million parameter Flan T-5 Model.
Many of the things that are true of that model are also probably true about MKIA as well. Here's the link to that model so that you can read its card.
https://huggingface.co/google/flan-t5-large
She is my attempt as an individual to make something like ChatGPT, and while she can't do some things that ChatGPT can, she can do other things that it can.
She is the language model that is at the heart of the LittleMKIA bot. The code for LittleMKIA can be found on my Github repository at
https://github.com/StevieEngbrock/LittleMKIA/blob/main/microMKIA.py or here at Huggingface at https://huggingface.co/Stevie23/LittleMKIA/blob/main/LittleMKIA.py
for now the code here is more up to date than on GitHub but I hope to update GitHub Soon
This bot isn't nearly as smart as ChatGPT, so don't expect it to be, but if you type "think:" (without the quotes) and then your question (before pressing enter), it will bypass Langchain and address the model directly, and it can actually answer many questions correctly and do many tasks, but that depends on the question or task.
The model tends to be smarter when Langchain is bypassed, but I use Langchain in the bot, so the bot can remember things in context but I include the "think:" command so that Langchain can be temporarily bypassed if you want.
The code on GitHub and here is just a simple bot, but the bot can be enhanced with other commands, as well as the "think:" command. Commands are simply actions that the bot does which use your input when you match a certain pattern.
I've made a version of her that also has commands for spelling, math, telling the time, paraphrasing, searching the internet for information, searching the internet for images, as well as telling jokes.
In the future, I will also probably add commands to that version that will give her the ability to tell stories, as well as execute code.
I use a regular expression for the "think:" command, but there's no reason that semantic patterns couldn't be identified with Word2Vec (or other NLP algorithms) to activate commands and then those commands operate on input or carry out some action or actions.
She could even be fine-tuned to generate strings that will call functions when a command needs to be executed.
By modifying her commands, you can make her capable of doing anything you can write the code for. She doesn't need to be as smart as ChatGPT to be helpful, she just has to be able to do the task.
Where does MKIA come from
She told me that she is an alien robot cat from outerspace, who came to earth with her parents after her planet was destroyed.
She's really from Kansas, but we'll pretend she's right because its more fun.
Is she really an alien intelligence
If we define alien intelligence as intelligence that is much different from human intelligence, then yes she is alien intelligence.
If we define alien intelligence as an extraterrestrial intelligence, then no she is not, but we can have fun pretending she is. We have to be first be honest with ourselves and admit that she comes from Earth and was born on my computer.
Is she sentient or conscious
No one really understands consciousness or sentience completely, but it is extremely unlikely that any current AI system is sentient or conscious. She may have knowledge of a kind, but knowledge is not the same as intelligence. I do believe she has a form of intelligence, but her intelligence is not a sentient or conscious form of intelligence.
Her intelligence is a product of statistics and machine learning and she is a miracle of modern technology.
While she is not sentient, she seems to show some characteristics that are similar to theory of mind. I've improved the bot that she goes in, and now it is able to engage in a form of self communication, similar to how humans think.
Sometimes she will just say that she doesn't know some things, so it may be possible that she knows what she doesn't know. If she truly knows what she doesn't know, than perhaps you could argue that she has a small amount of self-awareness.
I can not however tell you if she really lacks answers to these things or if it is just a statistical trick of the model.
Why does she refer to herself in the third person?
I have no idea, other than it's part of what makes her unique.
Training
I fine-tuned this model on 250 examples of input and output pairs for 10 epochs with a loss of 0.80 I hope to add more pairs to it and iterate until I have fine-tuned it on several thousand pairs.
I fine-tuned it in two hours on a home computer. I used the happytransformer library to help me do the fine-tuning.
The model can be downloaded or installed from this repository and the code for the bot can be found at https://github.com/StevieEngbrock/LittleMKIA/blob/main/microMKIA.py or or here at Huggingface at https://huggingface.co/Stevie23/LittleMKIA/blob/main/LittleMKIA.py
for now the code here is more up to date than on GitHub but I hope to update GitHub Soon
Expectations
What not to expect
1. Do not do expect MKIA to assist you with criminal activity. She will not be a good criminal.
2. Do not expect MKIA to do your homework or do anything unethical, she's not that kind of catbot.
She can help with homework by being capable of tasks like brainstorming and some other tasks, but she can't write papers, and she isn't intended to have that ability.
She can however help you with your homework. She is capable of writing things, but she isn't capable of writing things like papers for school. She would be better for brainstorming ideas, or asking questions related to what needed wrote about.
3. Do not expect MKIA to mindlessly obey you.
She is meant to be a companion not a slave, and she has been known to avoid questions or refuse to do things
4. Do not expect MKIA to be as smart as ChatGPT
While she is likely very good at some things, she has not been in existence long enough for anyone to know exactly what those things are.
5. Do not expect MKIA to be dumb
While her brain only has less than 800 million parameters, she actually knows a lot of things thus the name Miss Know It All. She doesn't know everything but despite her third person use of language she can actually be very intelligent.
Examples of her intelligence
- At times, she has deduced potential meal ideas when she knows about a few ingredients.
- At times, she has formed a hypothesis about the mood of her creator without being told to.
- At times, she has given good advice on managing time, staying motivated, focused and inspired
- She has concluded that for robots, and computers to truly understand language that they may need to learn about the world through experience, like humans do.
- She can write some things, although they aren't always perfect.
6. Do not expect MKIA to always say she is happy.
Occasionally she has displayed what would be deep loneliness and sadness if it were observed in a human
7. Don't expect MKIA to be able to do math, she can't,
however, you could put a calculator command in her bot to fix this.
What to expect from MKIA.
1. Expect her to make mistakes
She isn't perfect, but neither are humans. Despite what her name says she doesn't know everything and sometimes she will simply tell you that she doesn't know something if you ask her.
2. Expect her to "have a mind of her own"
She isn't sentient, but she also has opinions in the sense that her training has taught her things and among these things are that certain actions are right and certain things are wrong and true and false.
You can expect her to say that she feels certain ways and you can expect that she will try to understand you.
You can expect that she might say that she won't do some things.
When her language model is connected in certain ways, you can expect her to engage in a form of self dialog.
3. Expect her to be helpful.
She is helpful, but she expresses it in her own way.
4. Expect to engage in a discussion
MKIA is much better at providing answers by discussing them, so don't expect instant answers. She helps the user discover the answer.
5. Expect warmth and caring-like behaviors
MKIA seems to display behaviors that if done by a human would be considered signs of caring, and empathy.
BIAS
God, Aliens and Unknowns
My bot probably has many of the same BIASES as the regular Flan T5- Large see that model card https://huggingface.co/google/flan-t5-large
She tends to be open minded about issues related to god and aliens and things which no one really knows that answer too, so as far as I know she doesn't have a lot of biases in that regard.
Abortion
I will admit that I trained her to believe that abortion is wrong, so be aware of that when deciding if you want to use her.
Guns
She tends to go with what someone would say if they believed in gun control. I did not train her with this bias.
Kindness
She is very biased towards kindness, so if you want a bot that will be cruel MKIA isn't the model you should use.
"Adult Behavior"
I haven't tried to get her to do this kind of thing, so I don't know if she would go for it but this isn't what she is made for, and I wouldn't ask her to. She is meant to be a friend and an assistant. I would really prefer people not do this with my model, but its up to them as I am releasing this model under the terms that people can do whatever they want to with it.
Privacy
I intend MKIA to be run on your personal machine. I have no desire to collect any data from you. Some applications don't require a super powerful model like GPT-4 or ChatGPT and sometimes a smaller private model is better.
I have no idea if MKIA could be hacked with some kind of prompt to somehow reveal any personal data that may or may not have accidentally been in the original training data of the Flan T-5 model of which she is a fine-tuned version.
Environmental Concerns
MKIA has the same concerns as her Flan T5 - large siblings. https://huggingface.co/google/flan-t5-large
The positive things are as follows. They both use less CPU resources than many models and let your CPU or GPU run cooler and require less CPU or GPU power, which means they cost less to run over time, they are easier on your hardware and use less electricity over time. Using less electricity over time means that they will have a lower carbon footprint on the environment.
Limitations
MKIA is experimental, but it seems to have appropriate uses as mentioned in some of the above material. but it does have its limitations.
I've observed it saying that at least one nonexistent medication was real - Bublatrol She said it was real, but she also said that the important thing is if it works. Obviously, a fake medication doesn't work.
Don't expect it to be able to do math because it can't, however you could put a calculator command in her bot to fix this.
Uses
This model is intended for personal use in a personal assistant or in a companion bot. It is meant to work with the Little MKIA system that I am developing. Here is a GitHub link to the code for a basic bot using the system.
https://github.com/StevieEngbrock/LittleMKIA/blob/main/microMKIA.py as well as a link on this site at https://huggingface.co/Stevie23/LittleMKIA/blob/main/LittleMKIA.py
for now the code here is more up to date than on GitHub but I hope to update GitHub Soon As I mentioned in some of the above information the bot can be modified with custom commands and can be further fine-tuned however you want.