tale

a ~47m parameter llama-style model trained from scratch on tinystories. i made it because my little brother kept asking me for bedtime stories.

run it

ollama run navthings/tale

or use the gguf with llama.cpp:

llama-cli -m tale-q8_0.gguf -p "once upon a time "

or try it in your browser: https://navthings.github.io/playground/

files

file	notes
tale-f16.gguf	unquantized
tale-q8_0.gguf	half the size, basically the same

details

13 layers, 384 wide, 6 query heads, 2 kv heads, 384 token context. trained on apple silicon with a warmup + cosine lr schedule. this release has only done 2000 steps so its early.

code: https://github.com/navthings/tale