Upload 3 files

README.md

@@ -1,14 +1,15 @@
 ---
-title: Queryshield
-emoji: 🐠
-colorFrom: yellow
-colorTo: purple
+title: QueryShield
+emoji: 🛡️
+colorFrom: blue
+colorTo: indigo
 sdk: gradio
-sdk_version: 6.13.0
+sdk_version: 4.44.0
 app_file: app.py
 pinned: false
 license: mit
-short_description: Multilingual prompt optimizer — rewrites raw queries into ex
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# QueryShield — Multilingual Prompt Optimizer
+
+Fine-tuned Qwen2.5-1.5B for multilingual prompt optimization.

app.py

@@ -0,0 +1,187 @@
+import gradio as gr
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+MODEL_ID = "nickoo004/queryshield-1.5b"
+
+SYSTEM = (
+    "You are QueryShield, a multilingual prompt optimizer. "
+    "Given a raw user question, rewrite it into a detailed instruction "
+    "prompt for a downstream LLM expert. "
+    "User language: {in_lang}. Response language: {out_lang}. "
+    "Expert role: {role}."
+)
+
+LANGUAGES = [
+    "English",
+    "Uzbek",
+    "Russian",
+    "Kazakh",
+    "Karakalpak",
+]
+
+ROLES = [
+    "Senior Software Engineer",
+    "Medical Expert",
+    "Financial Analyst",
+    "Legal Advisor",
+    "Data Scientist",
+    "Cybersecurity Specialist",
+    "Aerospace Engineer",
+    "Agricultural Scientist",
+    "Experienced Educator",
+    "Automotive Engineer",
+    "Pharmaceutical Researcher",
+    "Manufacturing Expert",
+    "Business Strategist",
+    "Professional Writer",
+    "Project Manager",
+    "Support Specialist",
+    "HR Consultant",
+    "Environmental Scientist",
+    "Mathematician",
+    "UX Designer",
+    "Research Professor",
+    "Nutritionist",
+    "Real Estate Consultant",
+    "Supply Chain Manager",
+    "Mechanical Engineer",
+    "Electrical Engineer",
+    "Civil Engineer",
+    "Physics Researcher",
+    "Chemistry Expert",
+    "Biology Researcher",
+]
+
+EXAMPLES = [
+    ["hey how do i fix memory leak in my python app? its getting slower over time", "English", "English", "Senior Software Engineer"],
+    ["menga diabetni boshqarish uchun eng yaxshi ovqatlanish rejimini ayting, qon qandim yuqori", "Uzbek", "Uzbek", "Medical Expert"],
+    ["как мне улучшить производительность SQL запросов? таблица очень большая", "Russian", "Russian", "Data Scientist"],
+    ["бизнесімді қалай бастауға болады? капиталым аз, бірақ идеям бар", "Kazakh", "Kazakh", "Business Strategist"],
+    ["balalarımda matematika sabaqları qıyın bolıp atır, qanday úyretiw kerek?", "Karakalpak", "Karakalpak", "Experienced Educator"],
+    ["uyimda elektr toki kesib qoldi, qanday muammoni o'zim hal qila olaman?", "Uzbek", "Russian", "Electrical Engineer"],
+    ["менің фермамда топырақ сапасы нашар, не істеуім керек?", "Kazakh", "Uzbek", "Agricultural Scientist"],
+]
+
+print("Loading model (CPU — this may take a minute)...")
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_ID,
+    torch_dtype=torch.float32,   # CPU requires float32
+    device_map="cpu",
+    trust_remote_code=True,
+)
+model.eval()
+print("✅ Model loaded")
+
+
+def optimize(user_question, input_language, output_language, role, max_new_tokens=400):
+    if not user_question.strip():
+        return "⚠️ Please enter a question."
+
+    messages = [
+        {"role": "system", "content": SYSTEM.format(
+            in_lang=input_language,
+            out_lang=output_language,
+            role=role,
+        )},
+        {"role": "user", "content": user_question},
+    ]
+    text = tokenizer.apply_chat_template(
+        messages, tokenize=False, add_generation_prompt=True
+    )
+    inputs = tokenizer(
+        text, return_tensors="pt", truncation=True, max_length=512
+    )
+    with torch.no_grad():
+        output = model.generate(
+            **inputs,
+            max_new_tokens=max_new_tokens,
+            temperature=0.7,
+            do_sample=True,
+            repetition_penalty=1.1,
+            pad_token_id=tokenizer.eos_token_id,
+        )
+    new_tokens = output[0][inputs["input_ids"].shape[1]:]
+    return tokenizer.decode(new_tokens, skip_special_tokens=True).strip()
+
+
+# ── UI ─────────────────────────────────────────────────────────────────
+with gr.Blocks(theme=gr.themes.Soft(), title="QueryShield") as demo:
+
+    gr.Markdown("""
+# 🛡️ QueryShield — Multilingual Prompt Optimizer
+
+Fine-tuned **Qwen2.5-1.5B** that rewrites raw user queries into expert-level instruction prompts for downstream LLMs.
+
+Supports **5 languages**: English · Uzbek · Russian · Kazakh · Karakalpak  
+Supports **cross-lingual routing**: write in one language, get instructions for another.
+
+> ⚠️ Running on **CPU** — generation takes ~30–60 seconds. Please be patient.
+
+📦 [Dataset](https://huggingface.co/datasets/nickoo004/queryshield-multilingual) · 
+🤖 [Model](https://huggingface.co/nickoo004/queryshield-1.5b) · 
+📓 [Kaggle Demo](https://www.kaggle.com/code/nursultankoshekbaev/queryshield-1-5b)
+""")
+
+    with gr.Row():
+        with gr.Column(scale=1):
+            question = gr.Textbox(
+                label="Raw User Question",
+                placeholder="Type your messy, natural question here...",
+                lines=4,
+            )
+            with gr.Row():
+                input_lang = gr.Dropdown(
+                    choices=LANGUAGES,
+                    value="English",
+                    label="Input Language",
+                )
+                output_lang = gr.Dropdown(
+                    choices=LANGUAGES,
+                    value="English",
+                    label="Output Language",
+                )
+            role = gr.Dropdown(
+                choices=ROLES,
+                value="Senior Software Engineer",
+                label="Expert Role",
+            )
+            max_tokens = gr.Slider(
+                minimum=100,
+                maximum=600,
+                value=400,
+                step=50,
+                label="Max output tokens",
+            )
+            btn = gr.Button("✨ Optimize Prompt", variant="primary")
+
+        with gr.Column(scale=1):
+            output = gr.Textbox(
+                label="Optimized Prompt (instruction for downstream LLM)",
+                lines=18,
+                show_copy_button=True,
+            )
+
+    gr.Examples(
+        examples=EXAMPLES,
+        inputs=[question, input_lang, output_lang, role],
+        label="📌 Example queries (click to load)",
+        cache_examples=False,
+    )
+
+    btn.click(
+        fn=optimize,
+        inputs=[question, input_lang, output_lang, role, max_tokens],
+        outputs=output,
+    )
+
+    gr.Markdown("""
+---
+**How it works:**  
+QueryShield sits between the user and the main LLM. It takes a raw query and outputs a structured instruction prompt — including role, tone, format, edge cases, and language routing instructions.
+
+Built with ❤️ by [nickoo004](https://huggingface.co/nickoo004)
+""")
+
+demo.launch()

requirements.txt

@@ -0,0 +1,5 @@
+transformers==4.44.0
+torch==2.3.0
+gradio==4.44.0
+accelerate==0.33.0
+sentencepiece