f3nsmart/TinyLlama-MBTI-Interviewer-LoRA. v2.0

app.py

@@ -1,96 +1,35 @@
 import gradio as gr
-import torch
-from transformers import (
-    AutoTokenizer,
-    AutoModelForCausalLM,
-    AutoModelForSequenceClassification,
-    pipeline
-)
-from peft import PeftModel  # 👈 важно для LoRA адаптации
+import asyncio
+from core.utils import generate_first_question
+from core.mbti_analyzer import analyze_mbti
+from core.interviewer import generate_question
 
 # ===============================================================
-# 1️⃣ Настройки и модели
-# ===============================================================
-MBTI_MODEL = "f3nsmart/MBTIclassifier"
-INTERVIEWER_BASE = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
-INTERVIEWER_LORA = "f3nsmart/TinyLlama-MBTI-Interviewer-LoRA"
-
-# --- MBTI классификатор ---
-mbti_pipe = pipeline("text-classification", model=MBTI_MODEL, return_all_scores=True)
-
-# --- Интервьюер TinyLlama + LoRA ---
-print("🔄 Загрузка TinyLlama с адаптером LoRA...")
-tokenizer_llama = AutoTokenizer.from_pretrained(INTERVIEWER_LORA)
-base_model = AutoModelForCausalLM.from_pretrained(
-    INTERVIEWER_BASE,
-    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
-    device_map="auto"
-)
-model_llora = PeftModel.from_pretrained(base_model, INTERVIEWER_LORA)
-
-llm_pipe = pipeline(
-    "text-generation",
-    model=model_llora,
-    tokenizer=tokenizer_llama,
-    max_new_tokens=70,
-    temperature=0.7,
-    top_p=0.9,
-    device_map="auto"
-)
-
-# ===============================================================
-# 2️⃣ Вспомогательные функции
+# 3️⃣ Интерфейс Gradio
 # ===============================================================
-def clean_question(text: str) -> str:
-    text = text.strip().split("\n")[0].strip('"').strip("'")
-    bad_tokens = ["user:", "assistant:", "instruction", "interviewer", "system:"]
-    for bad in bad_tokens:
-        if bad.lower() in text.lower():
-            text = text.split(bad)[-1].strip()
-    if "?" not in text:
-        text = text.rstrip(".") + "?"
-    if len(text.split()) < 3:
-        return "What do you usually enjoy doing in your free time?"
-    return text.strip()
-
-def generate_first_question():
-    return "What do you usually enjoy doing in your free time?"
-
-def analyze_and_ask(user_text, prev_count):
+async def analyze_and_ask_async(user_text, prev_count, user_id="default_user"):
     if not user_text.strip():
         return "⚠️ Введите ответ.", "", prev_count
+
     try:
         n = int(prev_count.split("/")[0]) + 1
     except Exception:
         n = 1
     counter = f"{n}/30"
 
-    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]])
+    mbti_task = asyncio.create_task(analyze_mbti(user_text))
+    interviewer_task = asyncio.create_task(generate_question(user_id, user_text))
 
-    prompt = (
-        f"User said: '{user_text}'. "
-        "Generate one natural, open-ended question that starts with 'What', 'Why', 'How', or 'When'. "
-        "Avoid rephrasing or quoting the user's text. "
-        "Do NOT explain what you are doing or include any instructions. "
-        "Output only the question itself."
-    )
+    mbti_text, next_question = await asyncio.gather(mbti_task, interviewer_task)
+    return mbti_text, next_question, counter
 
-    raw = llm_pipe(prompt)[0]["generated_text"]
-    cleaned = clean_question(raw)
-    if not cleaned.startswith(("What", "Why", "How", "When")):
-        cleaned = "What motivates you to do the things you enjoy most?"
-    return mbti_text, cleaned, counter
 
-# ===============================================================
-# 3️⃣ Интерфейс Gradio
-# ===============================================================
 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=1):
             inp = gr.Textbox(
@@ -104,7 +43,7 @@ with gr.Blocks(theme=gr.themes.Soft(), title="MBTI Personality Interviewer") as
             interviewer_out = gr.Textbox(label="💬 Следующий вопрос от интервьюера", lines=3)
             progress = gr.Textbox(label="⏳ Прогресс", value="0/30")
 
-    btn.click(analyze_and_ask, inputs=[inp, progress], outputs=[mbti_out, interviewer_out, progress])
-    demo.load(lambda: ("", generate_first_question(), "0/30"), inputs=None, outputs=[mbti_out, interviewer_out, progress])
+    btn.click(analyze_and_ask_async, inputs=[inp, progress], outputs=[mbti_out, interviewer_out, progress])
+    demo.load(lambda: ("", generate_first_question(), "0/30"), None, [mbti_out, interviewer_out, progress])
 
-demo.launch()
+demo.queue(concurrency_count=2).launch()

core/interviewer.py

@@ -0,0 +1,49 @@
+import torch, asyncio
+from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
+from peft import PeftModel
+from core.utils import clean_question
+from core.memory import update_user_context, get_user_context, was_asked
+
+INTERVIEWER_BASE = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
+INTERVIEWER_LORA = "f3nsmart/TinyLlama-MBTI-Interviewer-LoRA"
+
+print("🔄 Loading interviewer (TinyLlama + LoRA)...")
+
+tokenizer = AutoTokenizer.from_pretrained(INTERVIEWER_LORA)
+base_model = AutoModelForCausalLM.from_pretrained(
+    INTERVIEWER_BASE,
+    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
+    device_map="auto"
+)
+model = PeftModel.from_pretrained(base_model, INTERVIEWER_LORA)
+
+llm_pipe = pipeline(
+    "text-generation",
+    model=model,
+    tokenizer=tokenizer,
+    max_new_tokens=70,
+    temperature=0.7,
+    top_p=0.9,
+    device_map="auto"
+)
+
+
+async def generate_question(user_id: str, user_text: str) -> str:
+    """Асинхронная генерация вопроса"""
+    history = get_user_context(user_id)
+    prev_qs = " | ".join(history["questions"][-5:])  # последние 5 вопросов
+
+    prompt = (
+        f"User said: '{user_text}'. Previous questions: {prev_qs or 'None'}. "
+        "Generate one natural, open-ended question starting with 'What', 'Why', 'How', or 'When'. "
+        "Avoid repeating or rephrasing previous questions. "
+        "Output only the question itself."
+    )
+    loop = asyncio.get_event_loop()
+    result = await loop.run_in_executor(None, lambda: llm_pipe(prompt)[0]["generated_text"])
+    cleaned = clean_question(result)
+
+    if was_asked(user_id, cleaned):
+        cleaned = "What new challenges have you faced recently?"
+    update_user_context(user_id, cleaned, user_text)
+    return cleaned

core/mbti_analyzer.py

@@ -0,0 +1,11 @@
+from transformers import pipeline
+from core.utils import format_mbti_output
+
+MBTI_MODEL = "f3nsmart/MBTIclassifier"
+mbti_pipe = pipeline("text-classification", model=MBTI_MODEL, return_all_scores=True)
+
+async def analyze_mbti(text: str) -> str:
+    """Асинхронный анализ MBTI"""
+    loop = __import__("asyncio").get_event_loop()
+    result = await loop.run_in_executor(None, mbti_pipe, text)
+    return format_mbti_output(result[0])

core/memory.py

@@ -0,0 +1,16 @@
+user_memory = {}
+
+def get_user_context(user_id: str):
+    """Возвращает историю вопросов и ответов для пользователя"""
+    return user_memory.get(user_id, {"questions": [], "answers": []})
+
+def update_user_context(user_id: str, question: str, answer: str):
+    ctx = user_memory.setdefault(user_id, {"questions": [], "answers": []})
+    ctx["questions"].append(question)
+    ctx["answers"].append(answer)
+    return ctx
+
+def was_asked(user_id: str, new_question: str) -> bool:
+    """Проверяет, повторялся ли вопрос"""
+    ctx = get_user_context(user_id)
+    return new_question.strip().lower() in [q.lower() for q in ctx["questions"]]

core/utils.py

@@ -0,0 +1,20 @@
+def clean_question(text: str) -> str:
+    text = text.strip().split("\n")[0].strip('"').strip("'")
+    bad_tokens = ["user:", "assistant:", "instruction", "interviewer", "system:"]
+    for bad in bad_tokens:
+        if bad.lower() in text.lower():
+            text = text.split(bad)[-1].strip()
+    if "?" not in text:
+        text = text.rstrip(".") + "?"
+    if len(text.split()) < 3:
+        return "What do you usually enjoy doing in your free time?"
+    return text.strip()
+
+
+def generate_first_question():
+    return "What do you usually enjoy doing in your free time?"
+
+
+def format_mbti_output(res):
+    res_sorted = sorted(res, key=lambda x: x["score"], reverse=True)
+    return "\n".join([f"{r['label']} → {r['score']:.3f}" for r in res_sorted[:3]])