Well I dont know where to begin… Last month I started on this project called Nuggt because I was fed-up with how all the autonomous agents out there require GPT-4 (at least 3 months ago) and GPT-4 is expensive and I didnt have no API keys at that time. So I wanted to create something with GPT-3.5 and thats when this whole Nuggt story started.
Long story short why stop there mate why not make it run on a open source model… sounds crazy (for me at least cuz I am no AI legend). So every time a new LLM model came out I tested it with nuggt by adjusting my initial prompt. They all failed because models like Vicuna were good in imitating not reasoning (as highlighted by the ORCA paper).
However, as some of you might have noticed, models trained coding for displayed some form of reasoning, at least that is what I noticed with StarCoder. Unfortunately, StarCoder was close but not good or consistent.
Today, I have finally found our winner Wizcoder-15B (4-bit quantised). Here is a demo for you. In this demo, the agent trains RandomForest on Titanic dataset and saves the ROC Curve.
You can find the github repo at: GitHub - Nuggt-dev/Nuggt: An Autonomous LLM Agent that runs on Wizcoder-15B
Do check it out and give me your feedback.