Upload app.py with huggingface_hub

app.py

@@ -474,9 +474,16 @@ if ON_HF:
         @spaces.GPU(duration=30)
         def _hf_gpu_generate(messages, max_new_tokens=150):
             model = _hf_model.to("cuda")
-            text = _hf_tok.apply_chat_template(
-                messages, tokenize=False, add_generation_prompt=True
-            )
+            try:
+                # Ask the model not to emit chain-of-thought, if it supports it
+                text = _hf_tok.apply_chat_template(
+                    messages, tokenize=False, add_generation_prompt=True,
+                    enable_thinking=False,
+                )
+            except TypeError:
+                text = _hf_tok.apply_chat_template(
+                    messages, tokenize=False, add_generation_prompt=True
+                )
             inputs = _hf_tok(text, return_tensors="pt").to("cuda")
             with torch.no_grad():
                 out = model.generate(
@@ -1200,31 +1207,41 @@ How do I feel right now?"""
         response = self.call_llm(prompt, system_prompt=system)
 
         if response:
+            # Strip any chain-of-thought blocks the model emits, then ASCII-clean
+            response = re.sub(r"<think>.*?</think>", " ", response, flags=re.DOTALL | re.IGNORECASE)
+            response = re.sub(r"</?think>", " ", response, flags=re.IGNORECASE)
             response = response.encode('ascii', errors='ignore').decode('ascii').strip()
-            # Try JSON parse (MiniCPM5 with format param returns JSON)
+
+            parsed = {}
+            art_id = None
+            lines = []
+
+            # Find a JSON object even if the model wrapped it in prose
+            m = re.search(r"\{.*\}", response, flags=re.DOTALL)
+            candidate = m.group(0) if m else response
             try:
-                parsed = json.loads(response)
-                lines = parsed.get("s", [])
-                if not lines:
-                    lines = parsed.get("message", [])
+                parsed = json.loads(candidate)
+                if not isinstance(parsed, dict):
+                    parsed = {}
+                lines = parsed.get("s") or parsed.get("message") or []
                 if not lines:
-                    # Handle t1, t2, t3... format
                     lines = [parsed.get(f"t{i}", "") for i in range(1, 7) if parsed.get(f"t{i}")]
                 if not lines:
                     lines = [parsed.get(f"l{i+1}", "") for i in range(4) if parsed.get(f"l{i+1}")]
-                
-                # Get art ID from LLM response
                 art_id = parsed.get("art", None)
                 if art_id is not None:
                     try:
                         art_id = int(art_id)
                     except (ValueError, TypeError):
                         art_id = None
-            except (json.JSONDecodeError, KeyError, TypeError):
-                # Fallback: pipe-separated
-                response = response.replace('\n', ' ').replace('\r', '')
-                response = response.strip('|').strip()
-                lines = [p.strip() for p in response.split("|") if p.strip()]
+            except (json.JSONDecodeError, KeyError, TypeError, AttributeError):
+                parsed = {}
+
+            # No usable JSON lines? Derive short lines from the prose itself.
+            if not lines:
+                text = re.sub(r"\s+", " ", response).strip()
+                frags = re.split(r"[.!?\n|]+", text)
+                lines = [f.strip() for f in frags if len(f.strip()) > 2][:6]
 
             if lines:
                 # Deduplicate and clean
@@ -1258,8 +1275,8 @@ How do I feel right now?"""
                         if rnd.random() < 0.3:  # 30% chance to move
                             dirs = ["up", "down", "left", "right"]
                             self.grid.move(rnd.choice(dirs))
-                    
-                    return tuple(lines)
+
+                    return tuple(lines[:4])
 
         # Fallback with context awareness
         self.agent_thinking = False