Llama love

- 5 mins

How we built “LLamaLove”: A fine-tune of Llama-2-13B on 50 Shades of Grey at the Custom LLM Hackathon

Our ‘llama love’ bot hackathon team: Rongfei, Lisa, Me (Chris), Era, and Jesse, pictured from left to right.

Our ‘llama love’ bot hackathon team: Rongfei, Lisa, Me (Chris), Era, and Jesse, pictured from left to right.

Update:

I’ve deployed the bot on poe.com! You can try it out here: https://poe.com/llamalove

Looks like together.ai has turned off hosting of the finetune. This makes sense since they charge per hour, not per request, so this burned through the entry $ they provided customers within about a week.

The day started strong. Adam D’Angelo, CEO of Quora kicked us off by sharing what Quora has been up to with building in the generative AI space - Poe.com

Adam

Poe

Many more sponsors jumped in, such as Robert Nishihara and Ameer Haj Ali from Anyscale, giving us hope and proof that fine-tuning Llama2 on specific tasks can help open source models perform better than GPT4! 🔥🔥

Anyscale

Sergey, the Llama 2 creator, jumped in to give a quick tech talk on how he made llama 2

Sergey

And then, Jamie from Together AI shared how hackers could use the together AI API to very easily and quickly fine tune LLama 7B on 100M tokens in < 30 minutes.

Together

My team had an observation. We saw that the existing LLMs out there have an opportunity for ♥️ LOVE ♥️. Today, when we ask LLMs to give us something 🍝saucy🍝, it gives us boring answers like ”I apologize, but as a large language model, I cannot…”…

Pitch

So we thought … what if we could have LLMs generate content that gives us back a lil’ more ♥️ love ♥️?

And thus, we fine tuned 2 models. The first was trained on the transcripts of the TV shows

  1. Too hot to handle
  2. Love Island
  3. Temptation Island
  4. Naked Attraction
  5. Love is blind

This model was meant to act like a chat model. We parsed the input until it was in a format prepared for together.ai to pick it up and train a model for us on it.

The result: Given the prompt, “Hey Baby”, we get:

Too hot to handle

“, you’re looking good! I’m a single woman, and I’m looking for a single man.”

Not bad.

Next we found a pdf of the every page of the book ‘50 shades of grey’ - a book known for the love it creates.

After massaging our data into a reasonable format to support our fine-tuning needs,

{"text": "Are you trying to completely emasculate me?"}
{"text": "This is probably the only place that I’ll be able to afford to pay."}
{"text": "Anastasia, I appreciate the thought. I do. But no."}
{"text": "Do you want to come in? I ask shyly."}
{"text": "I need to work, Anastasia, but I’ll be back this evening. What time?"}
{"text": " I ignore the unwelcome stab of disappointment. Why do I want to spend every single minute with this controlling sex god? Oh yes, I’ve fallen in love with him, and he can fly."}
{"text": "Thank you… for the more."}
{"text": "My pleasure, Anastasia. He kisses me, and I inhale his sexy Christian smell."}

Together AI gives us a brand new completion model that gives us the feel of 50 shades of grey:

With the prompt of “We were both lying there, holding hands”, and using our brand new fine tuned model to perform chat completion, we get:

With these two small demos, we showed the judges at the hackathon that it certainly is possible for a group of hackers to get up and running with a fine tuned custom LLM model within hours.

You can view our pitch deck here: https://pitch.com/public/afe488e7-9697-40a8-8f67-21b6c8ad5229/8e30d52d-e17a-4df3-8809-8e6d45e46b65

Our video of our presentation here:

Google photos link share

Throughout this day, I got to work with some amazing minds such as Era Qian, Rongfei Lu, Marco Mascorro, Lisa Liu, and Jesse Hu. We got a team picture together during the hacking:

teampic.jpeg

And finally, the judges selected us as the 2nd place winners! I was so happy to shake hands with Adam D’Angelo, Marco Mascorro, Taranjeet Singh and Jamie de Guerre from Together AI accepting my new clout that I worked hard for alongside with Era who stayed until the end with me!

winpic.jpeg

Super Special thanks to Jeremy Nixon - a man who worked so hard to organize this hackathon - and he even hacked on a project DURING the hackathon. I don’t know how he pulls it off.

Marcos brought his robotic dog that talks - Pictured above is Jeremy Nixon holding the microphone up so the dog will be projected loud enough.

Marcos brought his robotic dog that talks - Pictured above is Jeremy Nixon holding the microphone up so the dog will be projected loud enough.

Chris Settles

Chris Settles

Technologist, ML for Fraud @ Uber.