Refactor _clean_generated_text function to improve text processing and handling of response markers

app.py

@@ -72,15 +72,44 @@ def generate_prompt(instruction: str, input_text: str = "") -> str:
 
 
 def _clean_generated_text(generated: str, prompt: str) -> str:
-    generated = generated.strip()
-    # remove prompt if accidentally included
-    if generated.startswith(prompt):
-        generated = generated[len(prompt):].strip()
-    # remove common markers
-    for cut in ["### Instruction:", "### Response:", "ځواب:"]:
-        if generated.startswith(cut):
-            generated = generated[len(cut):].strip()
-    return generated
+    import re
+
+    g = generated.strip()
+
+    # If the model included a response marker, prefer the content after the last one
+    resp_token = "### Response:"
+    if resp_token in g:
+        g = g.split(resp_token)[-1].strip()
+
+    # Remove the original prompt (if present anywhere)
+    try:
+        if prompt and prompt.strip() and prompt in g:
+            g = g.replace(prompt, "").strip()
+    except Exception:
+        pass
+
+    # Remove common markers that may appear anywhere
+    for cut in ["### Instruction:", "### Response:", "Instruction:", "Response:", "ځواب:"]:
+        g = g.replace(cut, "").strip()
+
+    # If the prompt contained an instruction section, remove that instruction text
+    try:
+        if "### Instruction:" in prompt:
+            ins_part = prompt.split("### Instruction:", 1)[1]
+            # stop at response if present
+            if "### Response:" in ins_part:
+                ins_text = ins_part.split("### Response:", 1)[0].strip()
+            else:
+                ins_text = ins_part.strip()
+            if ins_text and ins_text in g:
+                g = g.replace(ins_text, "").strip()
+    except Exception:
+        pass
+
+    # Trim surrounding non-word characters and excessive whitespace
+    g = re.sub(r"^[\W_]+|[\W_]+$", "", g).strip()
+
+    return g
 
 
 def _detect_device() -> int: