Babble

A family of 5 tiny character-level GPTs trained from scratch on the same data (FineWeb-Edu + everyday conversations). They ramble. That's the point.

Model	Params	Layers	d_model	Context	Val loss
Micro	0.11M	2	64	64	1.699
Tiny	0.82M	4	128	128	1.388
Small	2.74M	6	192	256	1.264
Medium	10.78M	6	384	256	1.167
Bigish	19.47M	8	448	256	1.18

Each model lives in its own subfolder with its own README (model card), the PyTorch weights (.pt) and a GGUF (.gguf).

• Training / chat / export code: chiken_sandwich.py (one file does it all)
• Chat with all models at once: python chiken_sandwich.py chat
• Do not trust the output. They are toys.