just a dumb LLM i tried making as a starter AI/ML project, its like REALLY dumb so dont keep your hopes up, its actually even funny how dumb the responses can be so please try it for yourself once! trained on gcp using the free 300$ credits! its not trained from scratch but a qwen model fine trained over the finest datasets of dadjokes and reddit memes
@satvikhardat
on
The dumbest LLM out there
16m 57s logged
Hosted on Digital ocean, hoping this doesnt stop working, i have digital ocean credits for a few months lets see how it goes
@satvikhardat
on
The dumbest LLM out there
Everyone keeps trying to make smarter and smarter AI.
I thought it would be funny to go the other way and make a model that is dumb as hell but still kind of understands the format of a chatbot.
I also wanted to learn how fine-tuning actually works instead of just prompting an existing model and calling it a day.
So I took Qwen3 0.6B, gave it a bunch of dumb joke-style examples, and trained it to answer questions with fake explanations and fake confidence scores.
# HOW TO TEST
- Go to the attached huggingface link.
- Click restart this session
- wait upto 5-10 mins (MAX)
- and use the AI :D
(Please dont reject this ship i cannot host it 24/7 hugging face spaces are the most cost effective for me here, running a model is costly gng)
## What it does
You can ask it anything:
```
javascript?
why is water wet?
what is linux?
are we dumb?
And it replies with nonsense like:
javascript is a small disaster with excellent branding.
Confidence: 92 potatoes.
```
The goal is not accuracy. The goal is “this sounds like an LLM that got dropped on its head but is still trying its best.”
## Tech stack
- Qwen3 0.6B
- LoRA fine-tuning
- Unsloth
- PyTorch
- Transformers
- PEFT
- Gradio
- Hugging Face Spaces
**Dataset**
I used two CSV joke datasets:
```
dad_jokes.csv
shortjokes.csv
```
(from kaggle!)
I converted them into chat-style training examples where the user asks a random question and the assistant responds with dumb fake logic.
The training examples looked roughly like:
```
User: what is gravity?
Assistant: gravity happens because tiny invisible interns are panicking.
Confidence: 81 microwaved thoughts.
```
The raw joke data was honestly kind of cursed, so I also added filtering after generation to remove some weird/offensive lines.
## How it works
The app loads:
```
the base Qwen3 0.6B model
my LoRA adapter
a Gradio web UI
a small cleanup filter for bad generations
```
# Problems I ran into
This was more annoying than I expected.
### Some stuff that broke:
- Python 3.14 broke the training libraries
- I had to use Python 3.11
- the first dataset made the model too normal
- Qwen thinking mode kept wasting tokens
- some generated jokes were way too cursed
- GitHub Pages could not call my VM API because HTTPS was annoying
- deploying on Hugging Face Spaces ended up being simpler
@satvikhardat
on
The dumbest LLM out there
19h 36m 43s logged
I fine-tuned a tiny LLM to be stupid on purpose not “bad because it fails accidentally”, but bad in a funny way.
The whole point of DumbGPT is that you ask it something normal and it gives you the most confident nonsense answer possible.
Example:
User: what is gravity?
DumbGPT:
gravity happens because tiny invisible interns are panicking.
Confidence: 81 microwaved thoughts.
That is basically the entire project.
