Loading…

▾ 🌐 Cross-dataset accuracy by EXP base (Public × Syn v3 × Ext v1)

▾ Dataset summary (across all EXPs)

▾ Experiments table
Click a row to load that EXP and switch to its Tasks view.
💡 hover Question column to expand · Recent K column shows pass/fail dots for the last K experiments on each task (★ = currently selected EXP)
Compare with:  

▾ Question / 質問

▾ Context files (input data) / 入力データ

▾ Gold answer + script / 正解と解法スクリプト

Output (gold.csv) / 期待される出力

Script (gold_scripts/<task_id>.py) / 解法スクリプト


    

  

  

    
▾ Predictions (model outputs) / 各モデルの予測

    

      

        

          
selected EXP — prediction

          

          

        

        

          
comparison EXP — prediction

          

          

        

      

    

  

  
  
  

    
▾ Side-by-side dialogue (LINE-style) / 対話比較

    

      

        

          
selected EXP

          

        

        

          
comparison EXP

          

        

      

    

  






  

    
    
    
    
    
    Keyboard: ←/→ navigate · Esc back
  

  

  


  
▾ Question

    

  


  
▾ Context files (input data)

    

  


  
▾ Gold answer + script

    

      
Output (gold.csv)

      

      
Script (gold_scripts/<task_id>.py)

      

    

  


  
▾ Flow comparison (Gold vs Agents)

    

      

        Side-by-side: gold script's logical steps (left) vs each agent's
        tool-call distillation. Same color = matched pair (the agent
        reproduced that gold step). Gold steps with grey border = the agent
        did not perform that step. Agent steps with grey border = exploration
        or extra computation not in gold.
      

      

    

  


  
▾ Agent runs

    

      

        
Agent A

        

        

        
Prediction

        

        
Final answer

        

        
Tool trace

        

      

      

        
Agent B

        

        

        
Prediction

        

        
Final answer

        

        
Tool trace

        

      

    

  






  

    
▾ Local vs LB accuracy (all EXPs)

    

      

        Each point = one EXP. X = local accuracy (binary mean from this EDA's scoring).
        Y = leaderboard accuracy (manually entered below or in artifacts/eda/lb_scores.json).
        Diagonal = perfect agreement; below = LB worse than local; above = LB better.
        Hover for run_id. Click to open that EXP.
      

      

      
Color: v3 paired runs · Copilot Public-50 EXPs · grey = LB missing

    

  


  

    
▾ LB score input (per EXP)

    

      

        Edit values inline. Saved to browser (localStorage) and reflected in the plot above.
        Click Download lb_scores.json to get the merged file → commit to artifacts/eda/lb_scores.json for persistence.
      

      

        
run_idlabeldatasetlocal accLB accnote

        
      

      

        
        
      

    

  






  

    
▾ EXP overview

    

  

  

    
▾ Accuracy breakdown (this EXP)

    

  

  

    
▾ Agent A config + prompt

    

  

  

    
▾ Agent B config + prompt

    

  

  

    
▾ Tool catalog