app.py

import gradio as gr
import spaces
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
import re

_model = None
_tokenizer = None
MODEL_ID = "ogulcanaydogan/Turkish-LLM-14B-Instruct"
SYSTEM = "Sen yardimci bir Turkce asistansin. Acik, dogal ve anlasilir sekilde Turkce yanit ver."

def _load():
    global _model, _tokenizer
    if _model is None:
        _tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
        _model = AutoModelForCausalLM.from_pretrained(
            MODEL_ID, torch_dtype=torch.float16, device_map="auto"
        )
    return _model, _tokenizer

def _clean(text):
    cjk = re.search(r"[\u4e00-\u9fff\u3400-\u4dbf\u3040-\u309f\u30a0-\u30ff]", text)
    if cjk:
        text = text[:cjk.start()]
    text = text.rstrip()
    if text and text[-1] not in ".!?:;)]\"\'":
        last = max(text.rfind("."), text.rfind("!"), text.rfind("?"), text.rfind("\n"))
        if last > len(text) // 2:
            text = text[:last + 1]
    return text.strip()

@spaces.GPU(duration=120)
def respond(message, history):
    model, tok = _load()
    prompt = f"<|im_start|>system\n{SYSTEM}<|im_end|>\n"
    for u, a in (history or []):
        prompt += f"<|im_start|>user\n{u}<|im_end|>\n"
        prompt += f"<|im_start|>assistant\n{a}<|im_end|>\n"
    prompt += f"<|im_start|>user\n{message}<|im_end|>\n<|im_start|>assistant\n"
    ids = tok(prompt, return_tensors="pt").to(model.device)
    out = model.generate(**ids, max_new_tokens=512, temperature=0.7, top_p=0.9,
                         repetition_penalty=1.1, do_sample=True, pad_token_id=tok.eos_token_id)
    resp = tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=False)
    resp = resp.split("<|im_end|>")[0] if "<|im_end|>" in resp else resp
    resp = resp.split("\nuser")[0] if "\nuser" in resp.lower() else resp
    return _clean(resp)

CSS = """
.gradio-container { max-width: 800px !important; margin: auto; }
footer { display: none !important; }
"""

with gr.Blocks(css=CSS, title="Turkish LLM 14B") as demo:
    gr.Markdown("## Turkish LLM 14B Chat\nTurkce egitilmis 14 milyar parametreli acik kaynak dil modeli. Qwen2.5-14B-Instruct uzerine Turkce veri setiyle fine-tune edilmistir.\n\n**MMLU-TR:** 0.5977 **|** **XCOPA-TR:** 0.6400 **|** [Model](https://huggingface.co/ogulcanaydogan/Turkish-LLM-14B-Instruct) **|** [Koleksiyon](https://huggingface.co/collections/ogulcanaydogan/turkish-llm-family-69b303b4ef1c36caffca4e94)")
    chatbot = gr.ChatInterface(
        fn=respond,
        examples=[
            "Kuantum bilgisayarlar klasik bilgisayarlardan nasil farklidir?",
            "Turkiye\'nin UNESCO Dunya Mirasi listesindeki yerleri nelerdir?",
            "Python ile bir web scraper nasil yazilir?",
            "Iklim degisikliginin Akdeniz bolgesine etkileri nelerdir?",
        ],
        cache_examples=False,
    )
    gr.Markdown("---\n*Turkish LLM Family - 7B, 14B, 32B modelleriyle acik kaynak Turkce dil modeli ailesi.*")

demo.launch(ssr_mode=False)