New model. Qwen/Qwen2.5-1.5B-Instruct. v1.1

app.py

@@ -6,98 +6,128 @@ from transformers import (
     pipeline
 )
 
-# =========================================================
-# 1️⃣ Настройки моделей
-# =========================================================
+# ===============================================================
+# 1️⃣ Настройки и модели
+# ===============================================================
+
+# Fine-tuned MBTI Classifier (твоя модель)
 MBTI_MODEL = "f3nsmart/MBTIclassifier"
-INTERVIEWER_MODEL = "Phi-3-mini-128k-instruct"
+mbti_pipe = pipeline("text-classification", model=MBTI_MODEL, return_all_scores=True)
+
+# Модель-интервьюер
+INTERVIEWER_MODEL = "Qwen/Qwen2.5-1.5B-Instruct"
 
-interviewer_tok = AutoTokenizer.from_pretrained(INTERVIEWER_MODEL)
-interviewer_model = AutoModelForCausalLM.from_pretrained(INTERVIEWER_MODEL)
-interviewer_pipe = pipeline(
+tokenizer_qwen = AutoTokenizer.from_pretrained(INTERVIEWER_MODEL)
+model_qwen = AutoModelForCausalLM.from_pretrained(
+    INTERVIEWER_MODEL,
+    torch_dtype="auto",
+    device_map="auto"
+)
+
+llm_pipe = pipeline(
     "text-generation",
-    model=interviewer_model,
-    tokenizer=interviewer_tok,
-    max_new_tokens=80,
+    model=model_qwen,
+    tokenizer=tokenizer_qwen,
+    max_new_tokens=70,
     temperature=0.7,
-    top_p=0.9
+    top_p=0.9,
 )
 
-# =========================================================
-# 2️⃣ Функции
-# =========================================================
-def generate_first_question():
-    prompt = (
-        "Begin a friendly MBTI-style conversation. "
-        "Ask one simple, open-ended question about the person's interests or emotions. "
-        "Do not mention that you are an AI or interviewer. Output only the question text."
-    )
-    q = interviewer_pipe(prompt)[0]["generated_text"].strip()
-    return q.split("\n")[0] if "?" in q else "What makes you feel most fulfilled in your daily life?"
-
+# ===============================================================
+# 2️⃣ Вспомогательные функции
+# ===============================================================
 
-def clean_question(text):
-    """Удаляем мусор из вывода модели"""
+def clean_question(text: str) -> str:
+    """
+    Удаляет все инструкции и оставляет чистый вопрос.
+    """
     text = text.strip()
-    for bad in ["You are", "instruction", "Generate", "MBTI", "task", "assistant:"]:
+
+    # Берём только первую строку, если LLM вдруг вывела много
+    text = text.split("\n")[0]
+
+    # Иногда Qwen вставляет кавычки — убираем
+    text = text.strip('"').strip("'")
+
+    # Если модель вывела "User:" / "Assistant:" / "Instruction:" и т.п.
+    bad_tokens = ["user:", "assistant:", "instruction", "interviewer", "system:"]
+    for bad in bad_tokens:
         if bad.lower() in text.lower():
-            text = text.split(bad, 1)[-1]
-    text = text.replace(":", "").replace("'", "").strip()
+            text = text.split(bad)[-1].strip()
+
+    # Если вопрос не оканчивается знаком вопроса — добавляем
+    if "?" not in text:
+        text = text.rstrip(".") + "?"
+
+    # Мини-страховка от мусора
     if len(text.split()) < 3:
-        return None
-    return text
+        return "What do you usually enjoy doing in your free time?"
 
+    return text.strip()
 
+def generate_first_question():
+    """Первый вопрос фиксированный (без ожидания генерации)"""
+    return "What do you usually enjoy doing in your free time?"
 
 def analyze_and_ask(user_text, prev_count):
+    """
+    Основная логика: анализ MBTI + генерация нового вопроса.
+    """
     if not user_text.strip():
         return "⚠️ Введите ответ.", "", prev_count
 
+    # Прогресс
     try:
         n = int(prev_count.split("/")[0]) + 1
     except Exception:
         n = 1
     counter = f"{n}/30"
 
+    # Анализ MBTI
     res = mbti_pipe(user_text)[0]
     res_sorted = sorted(res, key=lambda x: x["score"], reverse=True)
     mbti_text = "\n".join([f"{r['label']} → {r['score']:.3f}" for r in res_sorted[:3]])
 
+    # Промпт для Qwen — чёткий, чтобы не возвращала инструкцию
     prompt = (
-        f"The user said: '{user_text}'. "
-        "Ask one short, open-ended follow-up question to learn more about their feelings, motivations, or habits. "
-        "Avoid yes/no questions. Respond only with the question."
+        f"User said: '{user_text}'.\n"
+        "Generate exactly one short, natural, open-ended question about personality, emotions, or preferences. "
+        "Avoid meta explanations, instructions, or introductions. "
+        "Output only the plain question text without quotes or notes."
     )
-    gen = interviewer_pipe(prompt)[0]["generated_text"].strip()
-    question = gen.split("\n")[0]
-    if not question.endswith("?"):
-        question += "?"
+
+    # Генерация нового вопроса
+    raw = llm_pipe(prompt)[0]["generated_text"]
+    question = clean_question(raw)
+
     return mbti_text, question, counter
 
-# =========================================================
+# ===============================================================
 # 3️⃣ Интерфейс Gradio
-# =========================================================
-with gr.Blocks(theme=gr.themes.Soft(), title="Adaptive MBTI Interviewer") as demo:
-    gr.Markdown("## 🧠 Adaptive MBTI Interviewer\nОпредели личностный тип и получи следующий вопрос от интервьюера.")
+# ===============================================================
+
+with gr.Blocks(theme=gr.themes.Soft(), title="MBTI Personality Interviewer") as demo:
+    gr.Markdown(
+        "## 🧠 MBTI Personality Interviewer\n"
+        "Определи личностный тип и получи следующий вопрос от интервьюера."
+    )
 
     with gr.Row():
-        with gr.Column(scale=2):
-            inp = gr.Textbox(label="Ваш ответ", placeholder="Например: I enjoy working with people and organizing events.", lines=4)
+        with gr.Column(scale=1):
+            inp = gr.Textbox(
+                label="Ваш ответ",
+                placeholder="Например: I enjoy working with people and organizing events.",
+                lines=4
+            )
             btn = gr.Button("Анализировать и задать новый вопрос", variant="primary")
+        with gr.Column(scale=1):
+            mbti_out = gr.Textbox(label="📊 Анализ MBTI", lines=4)
+            interviewer_out = gr.Textbox(label="💬 Следующий вопрос от интервьюера", lines=3)
+            progress = gr.Textbox(label="⏳ Прогресс", value="0/30")
 
-        with gr.Column(scale=2):
-            mbti_out = gr.Textbox(label="📊 Анализ MBTI", lines=5)
-            question_out = gr.Textbox(label="💬 Следующий вопрос от интервьюера", lines=3)
-            counter = gr.Textbox(label="Прогресс", value="0/30")
-
-    # Первый вопрос при запуске
-    demo.load(fn=generate_first_question, inputs=None, outputs=question_out)
+    btn.click(analyze_and_ask, inputs=[inp, progress], outputs=[mbti_out, interviewer_out, progress])
 
-    # Кнопка анализа и генерации
-    btn.click(fn=analyze_and_ask, inputs=[inp, counter], outputs=[mbti_out, question_out, counter])
+    # Автоматическая загрузка первого вопроса
+    demo.load(lambda: ("", generate_first_question(), "0/30"), inputs=None, outputs=[mbti_out, interviewer_out, progress])
 
-# =========================================================
-# 4️⃣ Запуск
-# =========================================================
-if __name__ == "__main__":
-    demo.launch()
+demo.launch()