MMLU PRO Benchmark

#3
by sevapru - opened

Hi, ran your model for a MMLU PRO benchmark this night (On Jetson Thor in 120W mode)

  MMLU-Pro Benchmark
  Model: GadflyII/Qwen3-Coder-Next-NVFP4
  Endpoint: http://localhost:8868/v1
  Subjects: 14
  Total questions: 12032
  Workers: 16
  Max tokens: 4096
  Subject                    Correct    Wrong   Accuracy
..
  biology                        646       71     90.1%
  business                       654      135     82.9%
  chemistry                      958      174     84.6%
  computer science               345       65     84.1%
  economics                      709      135     84.0%
  engineering                    629      340     64.9%
  health                         621      197     75.9%
  history                        261      120     68.5%
  law                            639      462     58.0%
  math                          1237      114     91.6%
  other                          701      223     75.9%
  philosophy                     368      131     73.7%
  physics                       1125      174     86.6%
  psychology                     644      154     80.7%
...
  TOTAL                         9537     2495    79.26%

Write me if you want to add the information to the model card

Thanks! I will add it to the card!

Did you also test the BF16 as a baseline?

No, I have only (lol) 128GB of VRAM, wouldn't fit without swap memory -> will be too painful to run whole bench without concurency and enough caching space

Sign up or log in to comment