[bug] LlamaHandler format about gorilla HOT 4 OPEN

Thank you for catching this! This is indeed an oversight on our end.

Currently, we are observing minor improvements in both the Llama3-8B and Llama3-70B models, which do not affect their positions on the leaderboard. We will verify the results and address this issue in the next release by updating the handler and adjusting the leaderboard scores accordingly. This update will ensure that the Llama handler's prompt template aligns with the official one.

from gorilla.

No problem! Awesome -- thanks for your great work on this benchmark.

I have another (mostly unrelated) question as well -- are the benchmark numbers reported when running with temperature=0? I was unable to reproduce the scores for Llama-3-8B-Instruct and noticed the default temperature setting was 0.7.

from gorilla.

No problem! Awesome -- thanks for your great work on this benchmark.

I have another (mostly unrelated) question as well -- are the benchmark numbers reported when running with temperature=0? I was unable to reproduce the scores for Llama-3-8B-Instruct and noticed the default temperature setting was 0.7.

We are using temperature = 0.7 for reported llama3 numbers. If you would like to provide more information on what is different, we can investigate why you are not able to reproduce.

from gorilla.

If we are sampling but not seeding, then results will naturally not be deterministic. There is some unavoidable non-determinism that vLLM introduces, but I imagine most of the discrepancy stems from sampling with non-zero temperature.

from gorilla.