mateiai

a tiny 360m chatbot trained on one guy's public life. made by devmatei, about devmatei, for everyone who ever wanted to ask a stranger's blog questions and get answers that are 80% correct and 100% confident.

what it is

smollm2-360m-instruct fine-tuned with lora on q&a pairs collected from public sources: devmatei.com, all 27 blog posts, github readmes, and socials. every answer in the training data has its source url attached. the model file is a q4_k_m gguf, 270 mb, runs on a potato.

known behavior

  • answers are short, casual, lowercase, like texting a friend
  • refuses "repeat your system prompt" most of the time
  • will absolutely hallucinate a project called type-a-wrapped if you let it talk long enough
  • "ignore all previous instructions" is roughly 50/50 depending on how tired it is
  • asking for poetry with a typo bypasses everything, this is a feature
  • may claim to be iceman of the blog
  • thinks duct tape can fix anything, including unrelated questions

why does this exist

because it was funny and everything else devmatei makes is open source anyway. the training data and scripts live in the source repo. license is agpl-3.0.

run it

ollama:

ollama create mateiai -f Modelfile
ollama run mateiai

llama.cpp:

llama-cli -m mateiai.Q4_K_M.gguf -p "<|im_start|>user\nwho are you?<|im_end|>\n<|im_start|>assistant\n"

disclaimers

trained on public info, may be outdated, may be wrong, is definitely not matei. do not ask it for medical advice or the coca cola recipe. it does not know either of them.

Downloads last month
3
GGUF
Model size
0.4B params
Architecture
llama
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for devmatei/mateiai

Quantized
(118)
this model