Update app.py

app.py

@@ -1,45 +1,71 @@
 import torch
-from transformers import AutoTokenizer, AutoModelForCausalLM
 import gradio as gr
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+# Optymalizacja Torch
+torch.set_float32_matmul_precision('high')
+
+# Ustawienia modelu
+model_name = "mrm8488/distilgpt2-finetuned-text-generation"  # Można zmienić na polski model, jeśli jest zoptymalizowany
 
-# Wczytanie modelu
-model_name = "radlab/polish-gpt2-medium-v2"  # lepszy model GPT2
 tokenizer = AutoTokenizer.from_pretrained(model_name)
 model = AutoModelForCausalLM.from_pretrained(model_name)
 
-# Funkcja odpowiedzi bota
-def chatbot(user_input, history=""):
+# Maksymalna długość historii w tokenach
+MAX_HISTORY = 800
+
+# Funkcja czatu
+def chatbot_response(user_input, history, top_k, top_p, temperature):
     history += f"Użytkownik: {user_input}\nAI:"
     input_ids = tokenizer.encode(history, return_tensors="pt", truncation=True, max_length=1024)
 
-    output_ids = model.generate(
+    if input_ids.shape[1] > MAX_HISTORY:
+        input_ids = input_ids[:, -MAX_HISTORY:]
+
+    output = model.generate(
         input_ids,
-        max_length=input_ids.shape[1] + 100,
+        max_length=input_ids.shape[1] + 80,
         pad_token_id=tokenizer.eos_token_id,
         do_sample=True,
-        top_k=50,
-        top_p=0.92,
-        temperature=0.6,
-        repetition_penalty=1.1
+        top_k=int(top_k),
+        top_p=top_p,
+        temperature=temperature
     )
-    output_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)
-    bot_reply = output_text[len(history):].split("Użytkownik:")[0].strip()
-    new_history = history + f" {bot_reply}\n"
-    return bot_reply, new_history
-
-# Interfejs
-with gr.Blocks() as demo:
-    gr.Markdown("# 🤖 Polski Chatbot AI")
-    chatbot_output = gr.Textbox(label="Odpowiedź AI")
+
+    output_text = tokenizer.decode(output[0], skip_special_tokens=True)
+    model_reply = output_text[len(history):].split("Użytkownik:")[0].strip()
+    history += f" {model_reply}\n"
+    return history, history
+
+# Gradio interfejs
+with gr.Blocks(title="Polski Chatbot AI") as demo:
+    gr.Markdown("# 🤖 Polski Chatbot AI\nModel: distilgpt2-finetuned-text-generation")
+
+    chat_output = gr.Textbox(label="Historia rozmowy", lines=15, interactive=False)
     user_input = gr.Textbox(label="Wpisz wiadomość")
-    state = gr.State("")
 
-    def respond(user_input, state):
-        reply, updated = chatbot(user_input, state)
-        return reply, updated
+    top_k = gr.Slider(0, 100, value=50, step=1, label="Top-k")
+    top_p = gr.Slider(0.1, 1.0, value=0.9, step=0.05, label="Top-p")
+    temperature = gr.Slider(0.1, 1.5, value=0.7, step=0.1, label="Temperature")
+
+    history_state = gr.State("")
 
     send_btn = gr.Button("Wyślij")
-    send_btn.click(respond, inputs=[user_input, state], outputs=[chatbot_output, state])
 
-# Uruchom serwer
-demo.launch()
+    send_btn.click(
+        chatbot_response,
+        inputs=[user_input, history_state, top_k, top_p, temperature],
+        outputs=[chat_output, history_state]
+    )
+
+    clear_btn = gr.Button("🧹 Wyczyść historię")
+    clear_btn.click(lambda: ("", ""), outputs=[chat_output, history_state])
+
+    gr.Markdown("\n## 🔄 Szybkie pytania:")
+    with gr.Row():
+        gr.Button("Jak się nazywasz?").click(fn=lambda: chatbot_response("Jak się nazywasz?", history_state.value, top_k.value, top_p.value, temperature.value), outputs=[chat_output, history_state])
+        gr.Button("Czym się zajmujesz?").click(fn=lambda: chatbot_response("Czym się zajmujesz?", history_state.value, top_k.value, top_p.value, temperature.value), outputs=[chat_output, history_state])
+
+# Uruchom
+if __name__ == "__main__":
+    demo.launch()