constion97/llm_embeddings

Introduction

These scripts are really about exploring embeddings, specifically input embeddings.

Print Layers

If we need to print out the layers of an LLM, just call (by default it'll choose Gemma 2B)

python print_layers.py

and for other models such as mistral

python print_layers.py --model "mistralai/Mistral-7B-v0.1"

or llama 7b

python print_layers.py --model "meta-llama/Llama-2-7b-hf"

larger models such as llama-2-70b chat

python print_layers.py --model "meta-llama/Llama-2-70b-chat-hf" 

Print Tokens

If we need to print out the tokens of an LLM, just call (by default it'll choose Gemma 2B) and the phrase "Who is Ada Lovelace?"

python print_layers.py

and for other models such as mistral

python print_tokens.py --tokenizer "mistralai/Mistral-7B-v0.1" --prompt "Who is Kitty Purry?"

python extract_embeddings.py --tokenizer "meta-llama/Meta-Llama-3-8b-Instruct" --model "meta-llama/Meta-Llama-3-8b-Instruct" --embeddings_file "./output/llama3_8b_embeddings_layer.pth" --dimensions 4096

python visualize_cosine_similarity.py --tokenizer "meta-llama/Meta-Llama-3-8b-Instruct" --model "meta-llama/Meta-Llama-3-8b-Instruct" --embeddings_file "./output/llama3_8b_embeddings_layer.pth" --dimensions 4096 --prompt "Sit Sat Mat Bat Hat Cat Nap Kit Kat Dog Fish Tree Math London Paris Rio Berlin Sydney Moscow Red Blue Green Black White for while print loop"