new protagonist model is coming along! almost feeling happy with her topology

rebirth has been fantastic so far, but oh man, they really knocked it out of the park with Cosmo Canyon.

biiiiiiig yawn for a tiny girl #cat

cat

we’re making chalupas (objectively the best taco bell menu item)

neural style transfer results on a picture of my best little boy. from left to right we have: The Persistence of Memory (Dali), The Scream (Munch), and The Starry Night (Van Gogh).

art cat, art cat, art cat.

@snats.xyz hey! my recent research was pretty similar to what your most recent post is. i was exploring layerwise pruning in LLMs with an interest in downstream NLP tasks. in order to determine which layers to prune, we looked at how similar the representations were at each layers final hidden state

evaluating pruned versions of Qwen 2.5 on BoolQ. after removing layers from its decoder, we heal with a small amount of fine tuning (QLoRA) and re-evaluate. honestly pretty cool how robust even tiny models are to layer removal for this kind of task.

still trying to get a cycleGAN implementation working. texture transfer is a little better, per the cat's fur...

i wrote a neural style transfer implementation; now you can see my starry night #cat

i'm writing a cycleGAN implementation to style transfer from e.g. Monet to arbitrary input pictures for a project. however, it just looks like i'm sticking my cat in a microwave

viridis is such a good color scheme for a heat map. we're looking at layer-block similarity on Llama3.1 1B (left) and Llama 3.1 70B (right) -- each row is normalized independently.

#llm #ai #nlp