ably.do/hft.py

import os
import torch
import torch.nn as nn
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer
from datasets import load_dataset
from PIL import Image
import re
import pytesseract
import docx2txt
import PyPDF2

from huggingface_hub import login

login(f"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")

def load_file_catalog(catalog_path):
    with open(catalog_path, 'r', encoding='utf-8') as file:
        return json.load(file)

def identify_legal_document(filename, file_catalog):
    return file_catalog.get(filename, f"")

# Funkcja do ekstrakcji tekstu z różnych typów plików
def extract_text_from_file(file_path):
    _, ext = os.path.splitext(file_path)
    ext = ext.lower()
    
    if ext in ['.txt', '.md']:
        with open(file_path, 'r', encoding='utf-8') as file:
            return file.read()
    elif ext == '.pdf':
        text = ""
        with open(file_path, 'rb') as file:
            reader = PyPDF2.PdfReader(file)
            for page in reader.pages:
                text += page.extract_text()
        return text
    elif ext in ['.doc', '.docx']:
        return docx2txt.process(file_path)
    elif ext in ['.jpg', '.jpeg', '.png', '.bmp', '.tiff']:
        return pytesseract.image_to_string(Image.open(file_path))
    else:
        return ""

# Przygotowanie danych
def prepare_dataset(directory, catalog_path):
    file_catalog = load_file_catalog(catalog_path)
    data = []
    for root, _, files in os.walk(directory):
        for file in files:
            file_path = os.path.join(root, file)
            text = extract_text_from_file(file_path)
            if text:
                # Sprawdzenie, czy plik znajduje się w katalogu
                doc_type = identify_legal_document(file, file_catalog)
                if doc_type != "Opracowanie własne":
                    # Przetwarzanie dla aktów prawnych
                    articles = re.split(r'(Art\.\s+\d+\.)', text)[1:]
                    for i in range(0, len(articles), 2):
                        if i + 1 < len(articles):
                            article_number = articles[i].strip()
                            article_content = articles[i + 1].strip()
                            data.append({
                                "text": f"{article_number} {article_content}",
                                "source": f"{doc_type}, {article_number}"
                            })
                else:
                    # Przetwarzanie dla zwykłych dokumentów
                    chunks = [text[i:i + 512] for i in range(0, len(text), 512)]
                    for chunk in chunks:
                        data.append({
                            "text": chunk,
                            "source": f""
                        })
    return data


# Tokenizacja danych
def tokenize_function(examples):
    inputs = tokenizer(examples["text"], truncation=True, padding="max_length", max_length=512)
    inputs["source"] = examples["source"]
    return inputs

# Dostosowany model
class CustomModel(AutoModelForCausalLM):
    def __init__(self, config):
        super().__init__(config)
        self.source_embedding = nn.Embedding(1000, config.hidden_size)  # Zakładamy maksymalnie 1000 różnych źródeł

    def forward(self, input_ids, attention_mask=None, labels=None, source=None):
        outputs = super().forward(input_ids, attention_mask=attention_mask, labels=labels)
        if source is not None:
            source_embeds = self.source_embedding(source)
            outputs.logits += source_embeds.unsqueeze(1)
        return outputs

# Dostosowany Trainer
class CustomTrainer(Trainer):
    def compute_loss(self, model, inputs, return_outputs=False):
        labels = inputs.pop("labels")
        source = inputs.pop("source")
        outputs = model(**inputs, labels=labels, source=source)
        loss = outputs.loss
        return (loss, outputs) if return_outputs else loss

# Przygotowanie modelu i tokenizera
model_name = "google/gemma-2-2b"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = CustomModel.from_pretrained(model_name)

# Przygotowanie datasetu
catalog_path = "file_catalog.json"
data = prepare_dataset("files", catalog_path)
dataset = load_dataset("dict", data=data)
tokenized_dataset = dataset.map(tokenize_function, batched=True, remove_columns=dataset["train"].column_names)

# Konfiguracja treningu
training_args = TrainingArguments(
    output_dir="./results",
    num_train_epochs=3,
    per_device_train_batch_size=4,
    save_steps=10_000,
    save_total_limit=2,
)

# Inicjalizacja Trainera
trainer = CustomTrainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset["train"],
)

# Trening modelu
trainer.train()

# Zapisanie modelu
trainer.save_model("./gemma2_finetuned")

# Funkcja do generowania odpowiedzi z cytowaniem
def generate_answer(question, model, tokenizer, dataset):
    inputs = tokenizer(question, return_tensors="pt")
    outputs = model.generate(**inputs, max_length=200, num_return_sequences=1, output_scores=True, return_dict_in_generate=True)
    
    answer = tokenizer.decode(outputs.sequences[0], skip_special_tokens=True)
    
    # Znajdź najbardziej prawdopodobne źródło
    source_probs = outputs.scores[-1][:, model.source_embedding.weight.shape[0]:]
    most_likely_source_idx = torch.argmax(source_probs).item()
    most_likely_source = dataset[most_likely_source_idx]['source']
    
    return f"{answer}\n\nŹródło: {most_likely_source}"

# Przykład użycia
question = "Ile dni urlopu przysługuje pracownikowi?"
answer = generate_answer(question, model, tokenizer, dataset)
print(answer)