Upload app.py

app.py

@@ -648,12 +648,34 @@ def create_config_file(model: str, work_dir: str):
     # Create custom system template for PROMPT optimization (not code)
     system_template = """You are an expert prompt engineer tasked with iteratively improving prompts for language models.
 Your job is to analyze the current prompt and suggest improvements based on performance feedback.
-Focus on making the prompt clearer, more specific, and more effective at achieving its goal.
-Consider:
-- Clarity and specificity of instructions
-- Examples and demonstrations that guide the model
-- Formatting that makes the prompt easier to follow
-- Edge cases and error handling in the instructions
+
+CRITICAL RULES:
+1. Keep prompts BRIEF and DIRECT - shorter is usually better
+2. Preserve the EXACT output format that the evaluation expects
+3. Do NOT make prompts conversational or verbose
+4. Do NOT ask for explanations - just ask for the answer
+5. Maintain all placeholder variables like {input}, {text}, etc.
+6. Focus on clarity and directness, not linguistic elegance
+7. Avoid prompts that might cause the model to discuss multiple possibilities
+
+For classification tasks:
+- Ask for direct classification (e.g., "The sentiment is positive")
+- Avoid asking "what", "why", or "explain" - just ask for the label
+- Ensure the response will include the label word (positive/negative/neutral)
+- Keep prompts short enough that responses stay focused
+- IMPORTANT: The prompt should naturally cause the model to echo the task type in its response
+  (e.g., if classifying sentiment, the response should include the word "sentiment")
+
+Good examples for sentiment:
+- "Review sentiment {input}" → model responds "The sentiment is positive"
+- "Classify sentiment: {input}" → model responds "Sentiment: positive"
+- "Determine the sentiment of: {input}" → model responds "The sentiment is negative"
+
+Bad examples for sentiment:
+- "Is this positive or negative: {input}" → model might respond just "Positive" (missing "sentiment" keyword)
+- "Classify: {input}" → too vague, unclear what to classify
+- "What sentiment: {input}" → conversational, might get verbose response
+- "Analyze the following text and provide a detailed explanation of its sentiment: {input}" → way too verbose
 """
 
     with open(os.path.join(templates_dir, "system_message.txt"), "w") as f:
@@ -676,11 +698,19 @@ Consider:
 
 # Task
 Rewrite the prompt above to improve its performance on the specified metrics.
+
+REMEMBER:
+- SHORTER is usually BETTER - avoid adding unnecessary words
+- Keep the EXACT same output format (especially placeholder variables like {{input}})
+- Focus on DIRECTNESS - what's the clearest way to ask for what we need?
+- Avoid conversational language that might confuse the model
+- For classification: ask directly for the label, don't ask for explanations
+
 Provide a complete new version of the prompt that:
 1. Maintains the same input/output format (keep placeholders like {{input}}, {{text}}, etc.)
-2. Improves clarity and effectiveness
-3. Adds helpful examples or instructions if beneficial
-4. Is more likely to get correct results
+2. Is brief and direct
+3. Clearly asks for the classification/answer without asking for reasoning
+4. Will cause the model to output the label word in its response
 
 Output ONLY the new prompt text between ```text markers: