Update src/streamlit_app.py

src/streamlit_app.py

@@ -412,52 +412,37 @@ def evaluate_dataset_with_gpt(subtask: str, df: pd.DataFrame, client=openai_clie
     sample_rows = df.head(3)[selected_cols].to_dict(orient="records")   # take 3 example rows
 
     prompt = f"""
-You are a data-validation assistant. Your job is to determine whether the dataset below is useful for the research subtask.
+You are a data‑validation assistant. Decide whether the dataset below is useful for the research subtask.
 
-===== TASK =====  
-Subtask: "{{subtask}}"
+===== TASK =====
+Subtask: "{subtask}"
 
-===== DATASET PREVIEW =====  
-Schema: {{schema}}  
-Example Rows: {{example_rows}}
+===== DATASET PREVIEW =====
+Schema (first {len(selected_cols)} columns):
+{json.dumps(column_info, indent=10)}
+Sample rows (10 max):
+{json.dumps(sample_rows, indent=10)}
 
 ===== OUTPUT INSTRUCTIONS (follow strictly) =====
+Case A – Relevant:
+• Write exactly two sentences, each no more than 30 words.
+• Summarize what the dataset contains and why it helps the subtask.
+• Do not mention column names or list individual rows.
+
+Case B – Not relevant:
+• Write one or two sentences, each no more than 30 words, **describing only what the dataset contains**.
+• Do **not** mention the subtask, relevance, suitability, limitations, or missing information (avoid phrases like “not related,” “does not focus,” “irrelevant,” etc.).
+• After the sentences, output the header **Additionally, here are some external resources you might find helpful:** on a new line. Format your output in markdown as:
+- [Name of Source](URL)
+• Then list 2–3 bullet points, each on its own line, starting with “- ” followed immediately by a URL likely to contain the needed data.
+• No additional commentary.
 
-Case A – Relevant:
-• Write exactly two sentences, each no more than 30 words.  
-• Sentence 1: summarize what the dataset contains.  
-• Sentence 2: explain why it helps answer the subtask.  
-• Do not mention specific column names or list individual rows.  
-• Do NOT generate any additional explanation or markdown formatting.
-
-Case B – Not Relevant:  
-• Write one or two sentences, each no more than 30 words, describing **only what the dataset contains**.  
-• Do **NOT** mention the subtask, usefulness, relevance, or missing information.  
-• Do **NOT** use words like “irrelevant,” “not related,” “not useful,” “not sufficient,” etc.  
-• After the sentence(s), output the exact header:
-
-**Additionally, here are some external resources you might find helpful:**
-
-• Then output 2–3 **real** resources in Markdown link format. Each must:
-   - Have a **real source name** (e.g., “MatWeb”, not “Name of Source”)
-   - Contain a **real, working URL** to a page or dataset related to the subtask  
-   - Be formatted exactly like: `- [Source Name](https://example.com)`
-
-• Do NOT use placeholder text like “Name of Source” or “URL”.  
-• Do NOT generate any commentary after the list.
-
-Example for Case B:
-
-The dataset contains technical specifications of commercial vehicles, such as engine types and dimensions.  
 
-**Additionally, here are some external resources you might find helpful:**  
-- [Polymer Property Database](https://polymerdatabase.com/)  
-- [MatWeb Materials Data Sheets](https://www.matweb.com/)  
-- [NIST Thermophysical Properties of Polymers](https://www.nist.gov/srd/nist-standard-reference-database-147)
 
+General rules:
+Plain text only — no code fences. Markdown link syntax (`[text](url)`) is allowed.
 """
 
-
     rsp = client.chat.completions.create(
         model="gpt-4o",
         messages=[{"role": "user", "content": prompt}],