kaikaidai commited on
Commit
da7b984
·
verified ·
1 Parent(s): 420bcc4

Add Selene-1-Mini

Browse files
Files changed (1) hide show
  1. app.py +13 -28
app.py CHANGED
@@ -686,7 +686,6 @@ with gr.Blocks(theme="default", css=CSS_STYLES) as demo:
686
  score3_description,
687
  score4_description,
688
  score5_description,
689
- is_first_game,
690
  ):
691
  # Build prompt data dictionary
692
  prompt_data = {
@@ -706,36 +705,24 @@ with gr.Blocks(theme="default", css=CSS_STYLES) as demo:
706
  if info.get("active", True)]
707
 
708
  # Define new models list
709
- new_models = ["Atla-8B-preview", "Flow-Judge-0.1", "SFR-LLaMA-3.1-70B-Judge"] # add "Flow-Judge-1.0" once ready
710
 
711
- if is_first_game:
712
- # For the first game, ensure Salesforce model is one of the models to catch up on votes
713
- salesforce_model = "SFR-LLaMA-3.1-70B-Judge"
714
- other_models = [m for m in active_models if m != salesforce_model]
 
715
  other_model = random.choice(other_models)
716
 
717
- # Randomly assign new model to either position A or B
718
  if random.random() < 0.5:
719
- model_a, model_b = salesforce_model, other_model
720
  else:
721
- model_a, model_b = other_model, salesforce_model
722
  else:
723
- # For subsequent games, new models appears 40% of the time
724
- if random.random() < 0.4:
725
- # Randomly choose between new models
726
- new_model = random.choice(new_models)
727
- other_models = [m for m in active_models if m not in new_models]
728
- other_model = random.choice(other_models)
729
-
730
- if random.random() < 0.5:
731
- model_a, model_b = new_model, other_model
732
- else:
733
- model_a, model_b = other_model, new_model
734
- else:
735
- # For other cases, exclude both Atla and Flow-Judge
736
- non_special_models = [m for m in active_models if m not in new_models]
737
- model1, model2 = random.sample(non_special_models, 2)
738
- model_a, model_b = (model1, model2) if random.random() < 0.5 else (model2, model1)
739
 
740
  # Get responses from models
741
  response_a = get_model_response(
@@ -811,7 +798,7 @@ with gr.Blocks(theme="default", css=CSS_STYLES) as demo:
811
 
812
  def handler(*args):
813
  nonlocal first_game
814
- result = submit_and_store(*args, first_game)
815
  first_game = False # Set to False after first submission
816
  return result
817
 
@@ -831,7 +818,6 @@ with gr.Blocks(theme="default", css=CSS_STYLES) as demo:
831
  score3_description,
832
  score4_description,
833
  score5_description,
834
- first_game_state, # Add first_game_state as input
835
  ],
836
  outputs=[
837
  score_a,
@@ -848,7 +834,6 @@ with gr.Blocks(theme="default", css=CSS_STYLES) as demo:
848
  model_name_b,
849
  send_btn,
850
  random_btn,
851
- first_game_state, # Add first_game_state as output
852
  ],
853
  )
854
 
 
686
  score3_description,
687
  score4_description,
688
  score5_description,
 
689
  ):
690
  # Build prompt data dictionary
691
  prompt_data = {
 
705
  if info.get("active", True)]
706
 
707
  # Define new models list
708
+ new_models = ["Atla Selene 1 Mini", "Flow-Judge-0.1", "SFR-LLaMA-3.1-70B-Judge"]
709
 
710
+ # New models appear 40% of the time
711
+ if random.random() < 0.4:
712
+ # Randomly choose between new models
713
+ new_model = random.choice(new_models)
714
+ other_models = [m for m in active_models if m not in new_models]
715
  other_model = random.choice(other_models)
716
 
 
717
  if random.random() < 0.5:
718
+ model_a, model_b = new_model, other_model
719
  else:
720
+ model_a, model_b = other_model, new_model
721
  else:
722
+ # For other cases, exclude new models
723
+ non_special_models = [m for m in active_models if m not in new_models]
724
+ model1, model2 = random.sample(non_special_models, 2)
725
+ model_a, model_b = (model1, model2) if random.random() < 0.5 else (model2, model1)
 
 
 
 
 
 
 
 
 
 
 
 
726
 
727
  # Get responses from models
728
  response_a = get_model_response(
 
798
 
799
  def handler(*args):
800
  nonlocal first_game
801
+ result = submit_and_store(*args)
802
  first_game = False # Set to False after first submission
803
  return result
804
 
 
818
  score3_description,
819
  score4_description,
820
  score5_description,
 
821
  ],
822
  outputs=[
823
  score_a,
 
834
  model_name_b,
835
  send_btn,
836
  random_btn,
 
837
  ],
838
  )
839