About the LCB evaluation
We used the officially recommended inference parameters but were only able to achieve around 60 points for LCB, not the 83 points reported on artificialanalysis.ai. Could you provide a more detailed explanation or the exact settings used?
Would like to know the details about your current evaluation settings.
Recommend using MiniMax Open Platform with temperature 1.0 and topp 0.95.
https://platform.minimax.io/docs/api-reference/text-intro
Temp=1.0
Topp=0.95
Topk=40
Max_tokens=128000
The parameters above are what we used during inference. We tried both local deployment testing with sglang and directly testing the API interface (last weekend), but the results were still relatively low. Also, our evaluation code is already fully aligned with the performance on other models.