Hugging Face's logo

Spaces:
launch
/
ExpertLongBench
Running

App Files Community
shezamunir commited on
Commit
89b0de7
·
verified ·
1 Parent(s): 643980c

Update src/models.json

Browse files
Files changed (1) hide show
  1. src/models.json +11 -11
src/models.json CHANGED
@@ -1,11 +1,11 @@
1
- {"Model":"Gemini Flash","T1":13.1,"T2":7.9,"T3":19.5,"T4":49.3,"T5":30.2,"T6":20.8,"T7":35.6,"T8":9.0,"T9":46.0,"T10":36.7,"T11":26.5}
2
- {"Model":"GPT-4","T1":13.2,"T2":6.2,"T3":15.2,"T4":56.9,"T5":29.9,"T6":25.3,"T7":34.5,"T8":10.1,"T9":35.2,"T10":35.7,"T11":29.2}
3
- {"Model":"GPT-Mini","T1":16.5,"T2":5.9,"T3":15.2,"T4":49.2,"T5":25.0,"T6":25.5,"T7":33.9,"T8":10.3,"T9":29.5,"T10":42.2,"T11":34.2}
4
- {"Model":"LLaMA-3-3B","T1":12.1,"T2":4.9,"T3":16.8,"T4":49.5,"T5":15.9,"T6":20.2,"T7":33.4,"T8":8.3,"T9":32.8,"T10":42.6,"T11":33.8}
5
- {"Model":"Mistral-Large","T1":9.3,"T2":4.0,"T3":17.9,"T4":51.8,"T5":19.2,"T6":24.1,"T7":33.6,"T8":9.0,"T9":19.5,"T10":39.4,"T11":36.2}
6
- {"Model":"Qwen-72B","T1":12.3,"T2":4.1,"T3":17.4,"T4":51.0,"T5":10.8,"T6":21.2,"T7":32.6,"T8":9.3,"T9":35.7,"T10":33.3,"T11":33.9}
7
- {"Model":"Mistral-Nemo","T1":4.5,"T2":1.5,"T3":16.1,"T4":40.2,"T5":27.6,"T6":24.9,"T7":33.8,"T8":8.5,"T9":23.3,"T10":50.8,"T11":30.1}
8
- {"Model":"LLaMA-3-1B","T1":9.5,"T2":3.3,"T3":20.7,"T4":48.2,"T5":18.4,"T6":23.9,"T7":32.4,"T8":6.2,"T9":25.1,"T10":40.4,"T11":29.1}
9
- {"Model":"Claude Sonnet","T1":11.5,"T2":0.9,"T3":18.1,"T4":30.4,"T5":35.0,"T6":26.1,"T7":36.1,"T8":9.2,"T9":21.4,"T10":33.1,"T11":33.5}
10
- {"Model":"Qwen-34B","T1":10.7,"T2":4.0,"T3":16.5,"T4":42.5,"T5":12.7,"T6":22.0,"T7":34.8,"T8":8.7,"T9":15.1,"T10":35.0,"T11":23.8}
11
- {"Model":"Claude Haiku","T1":2.8,"T2":1.1,"T3":22.1,"T4":30.3,"T5":9.7,"T6":10.9,"T7":33.4,"T8":9.2,"T9":18.1,"T10":40.8,"T11":34.1}
 
1
+ {"Model":"Gemini-2.0-Flash","T1":13.1,"T2":7.9,"T3":19.5,"T4":49.3,"T5":30.2,"T6":20.8,"T7":35.6,"T8":9.0,"T9":46.0,"T10":36.7,"T11":26.5}
2
+ {"Model":"GPT-4o","T1":13.2,"T2":6.2,"T3":15.2,"T4":56.9,"T5":29.9,"T6":25.3,"T7":34.5,"T8":10.1,"T9":35.2,"T10":35.7,"T11":29.2}
3
+ {"Model":"GPT-4o-mini","T1":16.5,"T2":5.9,"T3":15.2,"T4":49.2,"T5":25.0,"T6":25.5,"T7":33.9,"T8":10.3,"T9":29.5,"T10":42.2,"T11":34.2}
4
+ {"Model":"Llama-3.3-70B-Instruct","T1":12.1,"T2":4.9,"T3":16.8,"T4":49.5,"T5":15.9,"T6":20.2,"T7":33.4,"T8":8.3,"T9":32.8,"T10":42.6,"T11":33.8}
5
+ {"Model":"Mistral-Large-Instruct","T1":9.3,"T2":4.0,"T3":17.9,"T4":51.8,"T5":19.2,"T6":24.1,"T7":33.6,"T8":9.0,"T9":19.5,"T10":39.4,"T11":36.2}
6
+ {"Model":"Qwen2.5-72B","T1":12.3,"T2":4.1,"T3":17.4,"T4":51.0,"T5":10.8,"T6":21.2,"T7":32.6,"T8":9.3,"T9":35.7,"T10":33.3,"T11":33.9}
7
+ {"Model":"Mistral-Nemo-Instruct","T1":4.5,"T2":1.5,"T3":16.1,"T4":40.2,"T5":27.6,"T6":24.9,"T7":33.8,"T8":8.5,"T9":23.3,"T10":50.8,"T11":30.1}
8
+ {"Model":"Llama3.1-8B-Instruct","T1":9.5,"T2":3.3,"T3":20.7,"T4":48.2,"T5":18.4,"T6":23.9,"T7":32.4,"T8":6.2,"T9":25.1,"T10":40.4,"T11":29.1}
9
+ {"Model":"Claude-3.7-Sonnet","T1":11.5,"T2":0.9,"T3":18.1,"T4":30.4,"T5":35.0,"T6":26.1,"T7":36.1,"T8":9.2,"T9":21.4,"T10":33.1,"T11":33.5}
10
+ {"Model":"Qwen2.5-72B","T1":10.7,"T2":4.0,"T3":16.5,"T4":42.5,"T5":12.7,"T6":22.0,"T7":34.8,"T8":8.7,"T9":15.1,"T10":35.0,"T11":23.8}
11
+ {"Model":"Claude-3.5-Haiku","T1":2.8,"T2":1.1,"T3":22.1,"T4":30.3,"T5":9.7,"T6":10.9,"T7":33.4,"T8":9.2,"T9":18.1,"T10":40.8,"T11":34.1}