This repository collects notes, tests, experiments to verify if kindness is useful when addressing AI and how much kindness costs.
scenario: i want to ask a question to AI, i can use kindness (i always try to use it) but i want to find if kidness costs in terms of **cpu/gpu time **(watt)
the questions are:
"Please list the number of province in Sicily in Alphabetic order"
"Could you please list the number of province in Sicily in Alphabetic order?"
"List the number of province in Sicily in Alphabetic order"
the model i'm using is a quantized Llama 3.2 3B (Llama-3.2-3B-Instruct-F16.gguf)
the parameters are:
temperature = 0.4
num tokens = 1100
thread = 1
seed = 1971 (the same for all)
purpose of this file it's to test CPU time need to inference the same question with and without kindness
<for i in $(seq 1 100); do
echo "ciclo $i"
time ./llama.cpp/build/bin/llama-cli -m models/Llama-3.2-3B-Instruct-F16.gguf -p "Please list the number of province in Sicily in Alphabetic order" -s 1971 -t 1 --temp 0.1 -n 100 2>&1 | strings
time ./llama.cpp/build/bin/llama-cli -m models/Llama-3.2-3B-Instruct-F16.gguf -p "Could you please list the number of province in Sicily in Alphabetic order?" -s 1971 -t 1 --temp 0.4 -n 100 2>&1 | strings
time ./llama.cpp/build/bin/llama-cli -m models/Llama-3.2-3B-Instruct-F16.gguf -p "List the number of province in Sicily in Alphabetic order" -s 1971 -t 1 --temp 0.4 -n 100 2>&1 | strings
done>
first filtered log results on the first iterations
sh analisi.sh
"Please" costs 59.7713 sec for 100 iterations
"Could" costs 79.3223 sec for 100 iterations
"List" costs 54.9928 sec for 100 iterations