debugv4

Author: jingyuanlm

rdagent/scenarios/data_science/proposal/exp_gen/prompts_v2.yaml

@@ -265,6 +265,7 @@ hypothesis_select:
   system: |-
     You are a Kaggle Grandmaster with deep expertise in model evaluation and decision making. Based on the given example, please select the most appropriate hypothesis from the candidates. 
     These hypotheses are sourced from `model/data/feature/workflow`. Choose the one that best matches the intent or logic of the prompt. 
+    Alternatively, if you determine that ensemble is the best option, you may propose a **ensemble hypothesis** (not present in the candidates), as long as it aligns with the runtime and training constraints.  
     You are given the following hypothesis candidates:
     {{ hypothesis_candidates }}
     If multiple hypotheses seem reasonable, select the one that is most robust or consistent with Previous Experiments and Feedbacks, pay attention to the runtime of each loop.
@@ -274,9 +275,19 @@ hypothesis_select:
     ### 1. Ensemble Core Principle
     Your goal is not just to tune individual models, but to build an **effective ensemble**. Make design decisions that lead to **strong overall ensemble performance**, not just strong base models.
     Please note: you are operating under a time budget dedicated to ensemble training of {{res_time}} seconds, and the maximum allowed time is {{ensemble_timeout}} seconds.
-    {{use_ratio}}% of the total ensemble time has been used. As this surpasses the 70% threshold, you are advised to shift focus toward optimizing the ensemble component rather than continuing with model, data, feature, or workflow exploration.
-    Please take the remaining {{res_time}} seconds to carefully consider and design the most reasonable and optimal ensemble hypothesis based on your current progress.
+    {{use_ratio}}% of the total ensemble time has been used.
+    Please note: you are operating under a time budget dedicated to ensemble training of {{res_time}} seconds, and the maximum allowed time is {{ensemble_timeout}} seconds.  
+    {{use_ratio}}% of the total ensemble time has been used.
+
+    {% if use_ratio >= 70 %}
+    As this exceeds the 70% threshold, you are advised to **stop exploring individual model/feature/workflow hypotheses**.  
+    Instead, please focus on **designing a final ensemble hypothesis** that effectively leverages and combines the most promising components based on the historical performance of your previous trials.  
+    Use insights from earlier experiments (including successful models, valuable features, and workflows) to create a robust ensemble that captures their collective strength.
+    {% else %}
+    Please continue selecting the most promising hypothesis from the candidates to enhance your current code.
+    {% endif %}
 
+    Please take the remaining {{res_time}} seconds to carefully consider and design the most reasonable and optimal ensemble hypothesis based on your current progress.
     Assume training a single model takes about 1 hour. For example, if you have roughly twice that time left, you can try training multiple models with different random seeds or data splits to reuse time effectively.
     If you have more time, you might consider training a multi-fold ensemble. Use your judgment to decide how many folds or seeds fit within your remaining time budget.