talgat.py

# -*- coding: utf-8 -*-
import re, json, sys, subprocess
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from datasets import load_dataset
from tqdm import tqdm
#!pip install datasets==3.6.0
#
# ===== Параметры =====
BASE_MODEL   = "google/gemma-3-4b-it"
MODEL_PATH = "talgatzh/gemma-finetuned-model2"
OUTPUT_FILE  = "gemma_inference_results_from_multidomain_fixedzxcs555.jsonl"
MAX_NEW_TOKENS = 60
MAX_TEXTS    = 20  # увеличь для более стабильной метрики (>=200)

# ===== ROUGE (установим при необходимости) =====
try:
    import evaluate
except ImportError:
    subprocess.check_call([sys.executable, "-m", "pip", "install", "-q", "evaluate"])
    import evaluate

# ===== Модель и токенизатор =====
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_PATH,
    device_map="auto",
    torch_dtype=torch.float16,
    trust_remote_code=True
).eval()

# pad_token для стабильности (у Gemma pad = eos)
if tokenizer.pad_token_id is None:
    tokenizer.pad_token = tokenizer.eos_token

# ===== Утилиты =====
def is_kazakh(text: str) -> bool:
    return any(c in text.lower() for c in "қәөүңғұһі")

_SENT_SPLIT = re.compile(r'(?<=[\.\!\?…])\s+|\n+')
def lead_n(text: str, n=3) -> str:
    sents = [s.strip() for s in _SENT_SPLIT.split(text.strip()) if s.strip()]
    return " ".join(sents[:n])

def build_chat_prompt(text: str) -> str:
    instr = (
        "Мақсат: Экстрактивті қысқаша мазмұн.\n"
        "Ереже: Тек бастапқы мәтіндегі сөйлемдерді көшір. Өз сөзіңмен жазба. Синоним қолданба.\n"
        "Мәтіннен тек 2–3 ең маңызды сөйлемді таңда да, сол күйінде жаз.\n"
        "Формат: тек сөйлемдер, жаңа сөздер қоспа.\n\n"
        "Мәтін:\n"
        f"{text.strip()}\n\n"
        "Қысқаша мазмұн:"
    )
    messages = [{"role": "user", "content": instr}]
    # Правильный chat-template для Gemma-IT
    return tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

# ===== Данные =====
dataset = load_dataset("kz-transformers/multidomain-kazakh-dataset",
                       split="train", streaming=True)

INPUT_TEXTS = []
for ex in dataset:
    txt = (ex.get("text") or "").strip()
    if is_kazakh(txt) and len(txt.split()) > 20:
        INPUT_TEXTS.append(txt)
    if len(INPUT_TEXTS) >= MAX_TEXTS:
        break

print(f"✔ Отобрано {len(INPUT_TEXTS)} казахских текстов из multidomain")

# ===== Генерация =====
results, preds, refs = [], [], []

for text in tqdm(INPUT_TEXTS, desc="Generating summaries"):
    prompt_text = build_chat_prompt(text)

    toks = tokenizer(prompt_text, return_tensors="pt", truncation=True, max_length=2048)
    toks = {k: v.to(model.device) for k, v in toks.items()}

    with torch.no_grad():
        out = model.generate(
            **toks,
            max_new_tokens=MAX_NEW_TOKENS,
            do_sample=False,
            temperature=0.0,
            repetition_penalty=1.05,
            no_repeat_ngram_size=6,
            eos_token_id=tokenizer.eos_token_id,
            pad_token_id=tokenizer.pad_token_id,
            use_cache=True,
        )

    # === Берём ТОЛЬКО новые токены после входа ===
    input_len = toks["input_ids"].shape[1]
    gen_ids = out[0, input_len:]
    generated = tokenizer.decode(gen_ids, skip_special_tokens=True).strip()

    # Чистим возможные «утечки» ролей/маркеров
    for bad in ("model", "<start_of_turn>", "<end_of_turn>"):
        if generated.lower().startswith(bad):
            generated = generated[len(bad):].lstrip(": ").strip()
        generated = generated.replace(bad, "").strip()

    # Fallback: если пусто — берём первые 2–3 предложения исходника
    if not generated:
        generated = lead_n(text, n=3)

    reference = lead_n(text, n=3)

    results.append({"text": text, "summary": generated, "reference": reference})
    preds.append(generated)
    refs.append(reference)

# ===== Сохранение =====
with open(OUTPUT_FILE, "w", encoding="utf-8") as f:
    for r in results:
        f.write(json.dumps(r, ensure_ascii=False) + "\n")

print(f"✅ Сохранено {len(results)} суммаризаций → {OUTPUT_FILE}")

# ===== ROUGE к Lead-3 (прокси для быстрой диагностики) =====
rouge = evaluate.load("rouge")
scores = rouge.compute(predictions=preds, references=refs, use_stemmer=True)
scores_pct = {k: round(v * 100, 2) for k, v in scores.items()}
print("🔎 ROUGE vs Lead-3:")
for k in ("rouge1", "rouge2", "rougeL", "rougeLsum"):
    print(f"{k.upper()}: {scores_pct.get(k, 0)}%")

# ===== Быстрый дебаг первых 3 пар =====
for i in range(min(3, len(results))):
    print("\n--- SAMPLE", i+1, "---")
    print("PRED:", results[i]["summary"])
    print("REF :", results[i]["reference"])