Yeah we were definitely surprised! This was a small scale experiment; we uploaded our notebook for latency tests here if anyone is interested taking a look at the code / trying it their own finetuned models: https://github.com/eqian99/finetuning-experiment