参数调优：降低采样温度至 0.72（收紧 top_p=0.88，微调 repetition_penalty=1.18），提升文本连贯度与逻辑严谨性，消除离群采样

app.py

@@ -206,15 +206,15 @@ def background_generation_loop():
             model.to(device)
             inputs = tokenizer(prompt, return_tensors="pt").to(device)
 
-            add_server_log("SmolLM2-360M CPU 自回归推理中 (max_new_tokens=200)...")
+            add_server_log("SmolLM2-360M CPU 自回归推理中 (max_new_tokens=200, temp=0.72)...")
             with torch.no_grad():
                 output_ids = model.generate(
                     **inputs,
                     max_new_tokens=200,
                     do_sample=True,
-                    temperature=1.02,
-                    top_p=0.92,
-                    repetition_penalty=1.15,
+                    temperature=0.72,
+                    top_p=0.88,
+                    repetition_penalty=1.18,
                     bad_words_ids=BAD_WORDS_IDS,
                     pad_token_id=tokenizer.eos_token_id
                 )