haiku

README

HAIKUGEN

Documentation

Overview

haikugen generates JSON output for later eval cannot parallelize

data points (Xeon, PM: 2338/35224)

6 models x 10 message each: 1m53.604s, generation took between 8 and 0.04 seconds, output between 849 and 2 chars (14, or 25%). Around 3000% CPU usage.

6 models x 200 messages: 1200 replies in 34min or 0.58 haikus per second.