StupidLM

A small, useless GPT-style transformer. This model exists purely as a learning exercise — it generates mostly gibberish with the occasional grammatically correct sentence sprinkled in for false hope.

Do not use this for anything serious. You have been warned.

Model Details

Parameter	Value
Parameters	~23M
Architecture	Decoder-only Transformer
d_model	384
Heads	6
Layers	6
Context Length	256 tokens
Vocab Size	32,000

Checkpoints

File	Description
StupidLm_epoch_0.safetensors	After epoch 0 — barely learned anything
StupidLm_epoch_1.safetensors	After epoch 1 — slightly less clueless
StupidLm_epoch_2.safetensors	After epoch 2 — still useless, just confidently so

Usage

pip install torch safetensors tokenizers
python chat_safetensors.py

The chat script is fully self-contained — it includes the model architecture, loads the safetensors weights and tokenizer, and starts an interactive chat session. No external dependencies on the training repo.

Sample Output

You: hey
Assistant: Imagine there are two friends who are working together: Alice and Bob.
          They are working together each other, and there are some different
          ways to do this whole party.

Yeah. That's about as good as it gets.

Training

Trained using the JustAnEpoch training framework on conversational data for 3 epochs on a free Colab GPU. The name says it all.

Limitations

• It's tiny (~23M params). GPT-2 Small is ~5.4x bigger and still struggles.
• Trained for just 3 epochs on limited data.
• Generates repetitive, incoherent, and sometimes random CJK characters.
• Has no reasoning, no factual knowledge, no common sense.
• Basically a fancy random text generator with delusions of grandeur.