banking_model.py

import pandas as pd
import torch
from transformers import (
    AutoModelForCausalLM, 
    AutoTokenizer,
    TrainingArguments,
    Trainer,
    DataCollatorForLanguageModeling
)
from datasets import Dataset
import json
from pathlib import Path

class BankingModelTrainer:
    def __init__(
        self,
        base_model_name="meta-llama/Llama-2-13b-chat-hf",
        output_dir="./fine_tuned_model",
        max_length=512
    ):
        self.base_model_name = base_model_name
        self.output_dir = Path(output_dir)
        self.max_length = max_length
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        
        # تنظیمات مدل Llama-2
        model_config = {
            "device_map": "auto",
            "torch_dtype": torch.bfloat16,
            "low_cpu_mem_usage": True,
            "max_memory": {0: "10GB"},
            "load_in_8bit": True
        }

        # تنظیمات اولیه مدل و توکنایزر
        self.tokenizer = AutoTokenizer.from_pretrained(base_model_name)
        self.model = AutoModelForCausalLM.from_pretrained(
            base_model_name,
            **model_config
        )
        
    def prepare_data(self, data_path):
        # خواندن دیتا از فایل
        if data_path.endswith('.csv'):
            df = pd.read_csv(data_path)
        elif data_path.endswith('.json'):
            with open(data_path, 'r', encoding='utf-8') as f:
                data = json.load(f)
                df = pd.DataFrame(data)
        else:
            raise ValueError("فرمت فایل باید CSV یا JSON باشد")
        
        # پردازش و آماده‌سازی دیتا
        def prepare_examples(examples):
            conversations = []
            for q, a in zip(examples['question'], examples['answer']):
                # فرمت Llama-2 برای مکالمه
                conv = f"[INST] {q} [/INST] {a}"
                conversations.append(conv)
            
            # توکنایز کردن با تنظیمات Llama-2
            encodings = self.tokenizer(
                conversations,
                truncation=True,
                padding=True,
                max_length=self.max_length,
                return_tensors="pt"
            )
            
            return encodings
        
        dataset = Dataset.from_pandas(df)
        tokenized_dataset = dataset.map(
            prepare_examples,
            batched=True,
            remove_columns=dataset.column_names
        )
        
        return tokenized_dataset
    
    def train(self, dataset, epochs=3, batch_size=4):
        training_args = TrainingArguments(
            output_dir=str(self.output_dir),
            num_train_epochs=epochs,
            per_device_train_batch_size=batch_size,
            gradient_accumulation_steps=4,
            save_steps=500,
            logging_steps=100,
            learning_rate=2e-5,  # کاهش نرخ یادگیری برای Llama-2
            warmup_steps=100,
            fp16=True,  # فعال کردن fp16 برای Llama-2
            save_total_limit=2,
            logging_dir=str(self.output_dir / "logs"),
            gradient_checkpointing=True  # فعال کردن gradient checkpointing
        )
        
        data_collator = DataCollatorForLanguageModeling(
            tokenizer=self.tokenizer,
            mlm=False
        )
        
        trainer = Trainer(
            model=self.model,
            args=training_args,
            train_dataset=dataset,
            data_collator=data_collator
        )
        
        trainer.train()
        
        self.model.save_pretrained(self.output_dir)
        self.tokenizer.save_pretrained(self.output_dir)
        
    def generate_response(self, prompt):
        # فرمت Llama-2 برای پرامپت
        formatted_prompt = f"[INST] {prompt} [/INST]"
        inputs = self.tokenizer.encode(
            formatted_prompt, 
            return_tensors="pt"
        ).to(self.device)
        
        outputs = self.model.generate(
            inputs,
            max_length=self.max_length,
            num_return_sequences=1,
            temperature=0.7,
            top_p=0.9,
            do_sample=True,
            pad_token_id=self.tokenizer.eos_token_id,
            repetition_penalty=1.2  # اضافه کردن جریمه تکرار
        )
        
        response = self.tokenizer.decode(
            outputs[0], 
            skip_special_tokens=True
        )
        # حذف پرامپت از پاسخ
        response = response.replace(formatted_prompt, "").strip()
        return response

if __name__ == "__main__":
    trainer = BankingModelTrainer()
    dataset = trainer.prepare_data("banking_qa.json")
    trainer.train(dataset)
    response = trainer.generate_response("شرایط وام مسکن چیست؟")
    print(response)