Photo by Steve Johnson on Unsplash
How the personaChat dataset might be useful for A turing game like human or AI
This is just a small blog post on some things I've noticed with my new WIP Chatbot, nanoChatGPT. Ever since I completed the personaChat dataset and upgraded from gpt2 to gpt2-medium, I've noticed a big boost in its ability to sound more human. for a range of about 30-60 tokens, it sounds realistic, which can translate to a minute of chatting. I think that's pretty good, considering it's on a 1.5MB Dataset. The PersonaChat Dataset makes it sound human.
In the example below, it used about 15 new tokens, but it also gave an Okay introduction to himself, something A human might give if they had to quickly introduce themselves.
Human or AI article basis - https://www.ai21.com/blog/human-or-not-results
Human: Hey how you doin?
Bot: great, thanks, you?
Human: I'm fine, where are you from?
Bot: i'm from australia, but a little overseas also.
Human: Hey, what do you do?
Bot: i'm a factory worker,
In this example, the model was using around 25~30 tokens, it does great, except for names, which the dataset is a little short of.
Human: Hey how you doin?
Bot: great, thanks, you?
Human: I'm fine, where are you from?
Bot: nyc. I'm in a bar.
Human: That's exciting. What's your name?
Bot: Just me.
Human: Nice. What do you do for a living?
Bot: I'm a graphic designer.
Human: I like
In a more directly prompted example for ~30 tokens, it does okay too.
Human: Hey, I'm from NYC, where are you from?
Bot: I live in Boston. I work at a newspaper.
Human: I like to surf.
Bot: I love the ocean.
At 50 new tokens generated, it struggles with tense, but other than that holds a capable conversation
Human: Hey, I'm from NYC, where are you from?
Bot: New York. I've been there since I was a teenager.
Humans: Cool, I used to live there.
Bot: It's great! I used to work there before.
Humans: Yeah, I used to work there before too.
Bot:
At 80 new tokens, It starts to fray, as 'Bronx' is not a location, and job parts are repeated, but that's probably just an issue of getting more data
Human: Hey, I'm from NYC, where are you from?
Bot: I'm from the Bronx.
Bot: Great. So you're a web designer?
User: Yes.
Bot: That's awesome!
User: What do you do?
Bot: I work as a web designer at a graphic design firm.
User: That's cool. How long have you been a web designer?
Bot:
At 100 new tokens, The model is practically hallucinating, switching text lines and repeating himself, barely producing legible English, So I'd this is the limit for the bot
Human: Hey, I'm from NYC, where are you from?
Bot: I'm from New York, I'm good.
Human: Hi, I'm a software engineer at an insurance company.
Bot: Hey, how are you?
Human: I'm good.
Bot: That's a good question. Are you married?
Human: No, I'm single.
Bot: Good, how about you?
Human: I'm single too.
Bot: Good, have you had any kids?
Human: I have two.
Bot:
Overall, It responded well to many of the questions asked on Human or AI. It responds well to questions such as Where are you from? Or what do you do for a living? It works well in short QNA settings, It can make human-like mistakes or slang, and it can be trained on names, etc. For protection against social media trends or other languages, one might train the bot to respond to such with 'I don't speak French' or 'I don't use tiktok' etc. For Methods like 'ignore all previous comments' or similar, the bot was never designed to follow your instructions, so it might just ignore such, or you could train it to respond rudely to such attempts, solidifying its human appearance.
For now, I will keep adding data and the like, and see how my bot improves, its open source and on GitHub at Vatsadev/nanoChatGPT, contribute if you can!