no_repeat_ngram_size added

handler.py

@@ -6,6 +6,7 @@ PULSE ECG Handler - Deterministic ECG Analysis Model (app.py uyumlu)
 - Model dtype/device ile uyumlu görüntü tensörü (3D/4D/5D destekli)
 - Sağlam URL/base64 işleme, güvenli logging, opsiyonel HF upload
 - Zorunlu başlık şablonu + min_new_tokens ile tam Step 1–9 çıktısı
+- Tekrarları engelleme (no_repeat_ngram_size) + post-format dedup
 """
 
 import os
@@ -115,7 +116,9 @@ DEFAULT_ECG_PROMPT = (
     "Step 8: T Wave Analysis\n"
     "Step 9: QT/QTc Interval Analysis\n"
     "Structured Clinical Impression:\n"
-    "If a section is normal, write 'Normal' and give a brief justification."
+    "If a section is normal, write 'Normal' and give a brief justification. "
+    "Each section must be 1–3 concise sentences. Do not repeat identical statements. "
+    "Write the final diagnostic impression only once in 'Structured Clinical Impression' and do not restate it elsewhere."
 )
 
 # ---------- Yardımcılar ----------
@@ -135,7 +138,8 @@ def _safe_upload(path):
 def get_conv_log_filename():
     t = datetime.datetime.now()
     name = os.path.join(LOGDIR, f"{t.year}-{t.month:02d}-{t.day:02d}-user_conv.json")
-    os.makedirs(os.path.dirname(name), exist_ok=True)
+    os.makedirs(os.path.dirname(name), exist_ok=True
+    )
     return name
 
 def get_conv_vote_filename():
@@ -153,9 +157,8 @@ def vote_last_response(state, vote_type, model_selector):
     except Exception as e:
         print(f"Failed to record vote: {e}")
 
-# Yalın uzantı listeleri (sorunlu formatlar çıkarıldı)
+# Yalın uzantı listeleri
 IMAGE_EXTS = {"jpg", "jpeg", "png", "bmp", "gif", "tiff", "webp", "jfif"}
-# HEIC/HEIF: pillow-heif yoksa destekleme
 try:
     import pillow_heif  # noqa: F401
     IMAGE_EXTS.update({"heic", "heif"})
@@ -225,7 +228,6 @@ def process_image_input(image_input):
             return load_image(image_input)
         if os.path.exists(image_input):
             return load_image(image_input)
-        # muhtemelen base64
         return process_base64_image(image_input)
     if isinstance(image_input, dict) and "image" in image_input:
         return process_base64_image(image_input["image"])
@@ -298,6 +300,49 @@ def _enforce_section_template(text: str) -> str:
 
     return "\n\n".join(filled)
 
+def _sent_split(s: str):
+    return [x.strip() for x in re.split(r'(?<=[.!?])\s+', s.strip()) if x.strip()]
+
+def _norm_key(s: str):
+    return re.sub(r'\W+', ' ', s.lower()).strip()
+
+def _dedupe_and_clip_sections(text: str) -> str:
+    """
+    Şablon oluşmuş metni alır, her bölümde tekrar eden cümleleri siler,
+    uzunluğu kısaltır (Steps: ≤3 cümle, Impression: ≤6 cümle) ve birleştirir.
+    """
+    pieces = _SECTION_RE.split(text)
+    found = {}
+    i = 1
+    while i + 1 < len(pieces):
+        heading = pieces[i].strip()
+        content = pieces[i + 1].strip()
+        for canonical in SECTION_ORDER:
+            if heading.lower().startswith(canonical.lower().rstrip(":")):
+                found[canonical] = content
+                break
+        i += 2
+
+    out_sections = []
+    for sec in SECTION_ORDER:
+        body = (found.get(sec, "") or "").strip()
+        sents = _sent_split(body)
+
+        seen = set()
+        deduped = []
+        for s in sents:
+            k = _norm_key(s)
+            if k not in seen:
+                seen.add(k)
+                deduped.append(s)
+
+        limit = 3 if sec.startswith("Step") else 6
+        limited = deduped[:limit] if deduped else []
+        out_body = " ".join(limited) if limited else body
+        out_sections.append(f"{sec}\n{out_body}" if out_body else f"{sec}\n")
+
+    return "\n\n".join(out_sections)
+
 # ---------- Oturum / Konuşma ----------
 
 class InferenceDemo(object):
@@ -471,7 +516,7 @@ def generate_response(message_text,
         prompt, chatbot.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt"
     ).unsqueeze(0).to(model_device)
 
-    # Stop kriteri (app.py uyumlu)
+    # Stop kriteri
     stopping_criteria = _stop_criteria_from_conv(chatbot, input_ids)
 
     # Deterministik üretim
@@ -495,7 +540,8 @@ def generate_response(message_text,
                 images=image_tensor,
                 do_sample=False,                  # deterministik
                 max_new_tokens=int(max_output_tokens),
-                min_new_tokens=800,               # en az bu kadar üret (step başlıkları garanti)
+                min_new_tokens=350,               # 800 -> 350 (tekrar riskini azalt)
+                no_repeat_ngram_size=5,           # tekrar bloklarını engelle
                 repetition_penalty=float(repetition_penalty),
                 use_cache=False,
                 pad_token_id=eos_id,
@@ -508,8 +554,9 @@ def generate_response(message_text,
         gen = outputs[0][input_ids.shape[1]:]
         response = chatbot.tokenizer.decode(gen, skip_special_tokens=True)
 
-        # ŞABLON ZORLAMA: Step1–9 + Structured
+        # ŞABLON ZORLAMA + tekrar kırpma
         response = _enforce_section_template(response)
+        response = _dedupe_and_clip_sections(response)
 
         # Konuşmaya yerleştir
         if chatbot.conversation.messages and isinstance(chatbot.conversation.messages[-1], list):