Reserve answer space for long tokens; paramter test.

app.py

@@ -84,10 +84,16 @@ def evaluate_stream(msg, history, temperature, top_p):
     context += history[-1][0]
     context = context.replace(r'<br>', '')
 
+    # TODO: Avoid the tokens are too long.
+    CUTOFF = 160 
+    while tokenizer.encode(context) > CUTOFF:
+        # save 15 token size for the answer
+        context = context[15:]
+
     h = []
     print("History:", history)
     print("Context:", context)
-    for response, h in model.stream_chat(tokenizer, context, h, max_length=160, top_p=top_p, temperature=temperature):
+    for response, h in model.stream_chat(tokenizer, context, h, max_length=CUTOFF, top_p=top_p, temperature=temperature):
         history[-1][1] = response
         yield history, ""
 
@@ -108,7 +114,7 @@ with gr.Blocks() as demo:
         with gr.Column(scale=2):
             temp = gr.components.Slider(minimum=0, maximum=1.1, value=0.95, label="Temperature",
                 info="温度参数，越高的温度生成的内容越丰富，但是有可能出现语法问题。")
-            top_p = gr.components.Slider(minimum=0.5, maximum=1.0, value=0.99, label="Top-p",
+            top_p = gr.components.Slider(minimum=0.5, maximum=1.0, value=0.9, label="Top-p",
                 info="top-p参数，只输出前p>top-p的文字，越大生成的内容越丰富，但也可能出现语法问题。数字越小似乎上下文的衔接性越好。")
             #code = gr.Textbox(label="temp_output", info="解码器输出")
             #top_k = gr.components.Slider(minimum=1, maximum=200, step=1, value=25, label="Top k",