Spaces:
Running
What are Meta-Llama-3.1-Instruct "Turbo" models?
The leaderboard currently shows these models:
Meta Llama 3.1 8B Instruct TurboMeta Llama 3.1 70B Instruct TurboMeta Llama 3.1 4058B Instruct Turbo
However I do not see these "Turbo" names on the Hub. What do they mean?
cc
@clefourrier
Hey @m-ric , Judge Arena uses Together AI for inference of Llama (these quantised models are named as --- "Turbo") and the performance of the FP16 reference models seems(?) to be matched - some graphs on https://www.together.ai/blog/meta-llama-3-1
"Together Turbo achieves this performance while maintaining full accuracy compared to Meta’s reference implementation across all models. Llama-3.1-405B-Instruct-Turbo matches the accuracy of Meta reference models."
Would it make more sense to remove the "Turbo" from these names?
Hm I think it makes sense to keep the Turbo, but maybe we should add an url link from all model names to all versions used in the doc or table, wdyt?
Yes hopefully that prevents confusion - I'll add the url links in the leaderboard tab