Upload 3 files

README.md

@@ -0,0 +1,26 @@
+---
+title: QM-4B Chat
+emoji: 💬
+colorFrom: blue
+colorTo: purple
+sdk: gradio
+sdk_version: 5.9.1
+app_file: app.py
+pinned: false
+python_version: 3.12
+---
+
+# Къарачай-Малкъар чат-бот / Карачаево-балкарский чат-бот
+
+Чат-бот на основе модели [TSjB/QM-4B](https://huggingface.co/TSjB/QM-4B) для карачаево-балкарского языка.
+
+## Возможности
+
+- Генерация текста на карачаево-балкарском языке
+- Ответы на вопросы
+- Поддержка русского и английского языков
+
+## Разработчики
+
+- [Богдан Теунаев](https://t.me/bogdan_tewunalany)
+- [Али Берберов](https://t.me/ali_berberov)

app.py

@@ -0,0 +1,135 @@
+# Карачаево-балкарский чат-бот на основе TSjB/QM-4B
+import gradio as gr
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from datetime import datetime
+import pytz
+
+# Constants
+MODEL_LLM_PATH = 'TSjB/QM-4B'
+DEFAULT_SYSTEM_PROMPT = "Сен къарачай-малкъар тилде болушлукъчуса. Соруўлагъа къысха, тюз эм ачыкъ джуўабла бер."
+
+TZ = pytz.timezone('Europe/Moscow')
+
+# Load model
+print(f"Loading model {MODEL_LLM_PATH}...")
+tokenizer_llm = AutoTokenizer.from_pretrained(MODEL_LLM_PATH, trust_remote_code=True)
+model_llm = AutoModelForCausalLM.from_pretrained(
+    MODEL_LLM_PATH,
+    trust_remote_code=True,
+    torch_dtype=torch.bfloat16,
+    device_map="auto",
+)
+model_llm.eval()
+print("Model loaded!")
+
+
+def chat_llm(message, history, system_prompt, temperature, max_tokens, top_p, repetition_penalty):
+    """
+    Chat function for LLM model TSjB/QM-4B
+    """
+    if not message.strip():
+        return ""
+
+    # Build messages from history
+    messages = []
+    if system_prompt and system_prompt.strip():
+        messages.append({"role": "system", "content": system_prompt.strip()})
+
+    # Add history
+    for msg in history:
+        messages.append({"role": msg["role"], "content": msg["content"]})
+
+    # Add current message
+    messages.append({"role": "user", "content": message})
+
+    print(f'Chat input: {message} - Time: {datetime.now(tz=TZ)}')
+
+    # Apply chat template
+    try:
+        try:
+            text = tokenizer_llm.apply_chat_template(
+                messages,
+                tokenize=False,
+                add_generation_prompt=True,
+                enable_thinking=False
+            )
+        except TypeError:
+            text = tokenizer_llm.apply_chat_template(
+                messages,
+                tokenize=False,
+                add_generation_prompt=True
+            )
+    except Exception as e:
+        print(f"Error applying chat template: {e}")
+        text = message
+
+    # Tokenize
+    inputs = tokenizer_llm(
+        text,
+        return_tensors="pt",
+        truncation=True,
+        max_length=4096
+    ).to(model_llm.device)
+
+    if 'token_type_ids' in inputs:
+        inputs.pop('token_type_ids')
+
+    # Generate
+    with torch.no_grad():
+        outputs = model_llm.generate(
+            **inputs,
+            max_new_tokens=int(max_tokens),
+            temperature=float(temperature) if float(temperature) > 0 else 0.01,
+            top_p=float(top_p),
+            do_sample=True,
+            repetition_penalty=float(repetition_penalty),
+            pad_token_id=tokenizer_llm.pad_token_id,
+            eos_token_id=tokenizer_llm.eos_token_id,
+        )
+
+    # Decode only new tokens
+    response = tokenizer_llm.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
+
+    print(f'Chat output: {response[:100]}... - Time: {datetime.now(tz=TZ)}')
+
+    return response
+
+
+# UI
+with gr.Blocks(title="Къарачай-Малкъар ушакъ-бот", theme=gr.themes.Soft()) as demo:
+    gr.Markdown("# 💬 Къарачай-Малкъар ушакъ-бот")
+    gr.Markdown("Чат-бот на основе модели [TSjB/QM-4B](https://huggingface.co/TSjB/QM-4B) для карачаево-балкарского языка.")
+
+    with gr.Accordion("Настройки / Settings", open=False):
+        chat_system_prompt = gr.Textbox(
+            lines=2,
+            value=DEFAULT_SYSTEM_PROMPT,
+            label="Системаны промпту / System prompt"
+        )
+
+        with gr.Row():
+            chat_temperature = gr.Slider(minimum=0.0, maximum=2.0, value=0.7, step=0.1, label="Temperature")
+            chat_max_tokens = gr.Slider(minimum=16, maximum=2048, value=256, step=16, label="Max tokens")
+        with gr.Row():
+            chat_top_p = gr.Slider(minimum=0.1, maximum=1.0, value=0.9, step=0.05, label="Top-p")
+            chat_repetition_penalty = gr.Slider(minimum=1.0, maximum=2.0, value=1.2, step=0.1, label="Repetition penalty")
+
+    gr.ChatInterface(
+        chat_llm,
+        additional_inputs=[chat_system_prompt, chat_temperature, chat_max_tokens, chat_top_p, chat_repetition_penalty],
+        examples=[
+            ["Салам! Сен кимсе?"],
+            ["Къарачай-Малкъар тилни юсюнден айт."],
+            ["Минги Таў деген не болгъанды?"],
+            ["Привет! Расскажи о себе на русском."],
+        ],
+    )
+
+    gr.Markdown("""
+---
+**Разработчики:** [Богдан Теунаев](https://t.me/bogdan_tewunalany), [Али Берберов](https://t.me/ali_berberov)
+    """)
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,7 @@
+transformers==4.47.1
+tokenizers==0.21.0
+accelerate
+huggingface_hub==0.27.1
+torch
+gradio==5.9.1
+pytz