112 lines
4.1 KiB
Python
112 lines
4.1 KiB
Python
import os
|
|
import re
|
|
import torch
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer, DataCollatorForLanguageModeling
|
|
from datasets import Dataset
|
|
|
|
# Konfiguracja
|
|
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
|
MODEL_NAME = "gpt2"
|
|
SPECIAL_TOKENS = ["[CITATION_START]", "[CITATION_END]"]
|
|
TEXT_FILE_PATH = "./docs/kodekspracy.txt" # Zmień na właściwą ścieżkę
|
|
|
|
def prepare_dataset_from_file(file_path):
|
|
with open(file_path, 'r', encoding='utf-8') as f:
|
|
text = f.read()
|
|
|
|
# Wydziel artykuły za pomocą wyrażenia regularnego
|
|
articles = re.findall(r'Art\.\s*\d+[a-z]*\..*?(?=\s*Art\.\s*\d+[a-z]*\.|\Z)', text, flags=re.DOTALL)
|
|
|
|
formatted_articles = []
|
|
for article in articles:
|
|
# Usuń zbędne białe znaki
|
|
article = ' '.join(article.strip().split())
|
|
|
|
# Wydziel numer artykułu i treść
|
|
art_match = re.match(r'Art\.\s*(\d+[a-z]*)\.?\s*(.*)', article, re.DOTALL)
|
|
if art_match:
|
|
art_number = art_match.group(1)
|
|
art_text = art_match.group(2)
|
|
|
|
# Podziel na paragrafy, jeśli istnieją
|
|
paragraphs = re.split(r'(§\s*\d+\.)', art_text)
|
|
if len(paragraphs) > 1:
|
|
formatted_paragraphs = []
|
|
for i in range(1, len(paragraphs), 2):
|
|
para_num = paragraphs[i].strip()
|
|
para_text = paragraphs[i+1].strip()
|
|
formatted_paragraphs.append(f"{para_num} {para_text}")
|
|
formatted = f"[CITATION_START] Kodeks Pracy, Art. {art_number} [CITATION_END]\n" + "\n".join(formatted_paragraphs)
|
|
else:
|
|
formatted = f"[CITATION_START] Kodeks Pracy, Art. {art_number} [CITATION_END] {art_text}"
|
|
|
|
formatted_articles.append({"text": formatted})
|
|
|
|
return formatted_articles
|
|
|
|
def main():
|
|
# Inicjalizacja tokenizera
|
|
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
|
tokenizer.add_special_tokens({"additional_special_tokens": SPECIAL_TOKENS})
|
|
tokenizer.pad_token = tokenizer.eos_token
|
|
|
|
# Przygotowanie danych
|
|
data = prepare_dataset_from_file(TEXT_FILE_PATH)
|
|
dataset = Dataset.from_dict({"text": [d["text"] for d in data]})
|
|
|
|
# Tokenizacja
|
|
def tokenize_function(examples):
|
|
tokenized = tokenizer(
|
|
examples["text"],
|
|
truncation=True,
|
|
padding="max_length",
|
|
max_length=2048, # Zwiększono dla dłuższych artykułów
|
|
return_tensors="pt"
|
|
)
|
|
tokenized["labels"] = tokenized["input_ids"].clone()
|
|
return tokenized
|
|
|
|
tokenized_dataset = dataset.map(tokenize_function, batched=True, remove_columns=dataset.column_names)
|
|
|
|
# Model i data collator
|
|
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME)
|
|
model.resize_token_embeddings(len(tokenizer))
|
|
|
|
data_collator = DataCollatorForLanguageModeling(
|
|
tokenizer=tokenizer,
|
|
mlm=False
|
|
)
|
|
|
|
# Konfiguracja treningu
|
|
training_args = TrainingArguments(
|
|
output_dir="./results",
|
|
num_train_epochs=15, # Zwiększono liczbę epok
|
|
per_device_train_batch_size=4, # Zwiększono rozmiar batcha
|
|
learning_rate=2e-5, # Zmniejszono learning rate
|
|
weight_decay=0.01, # Dodano weight decay
|
|
logging_steps=10,
|
|
save_steps=500, # Dodano zapisywanie modelu co 500 kroków
|
|
eval_steps=500, # Dodano ewaluację co 500 kroków
|
|
evaluation_strategy="steps",
|
|
load_best_model_at_end=True, # Ładowanie najlepszego modelu na końcu
|
|
report_to="none",
|
|
save_total_limit=2, # Ograniczenie liczby zapisywanych checkpointów
|
|
)
|
|
|
|
# Trainer
|
|
trainer = Trainer(
|
|
model=model,
|
|
args=training_args,
|
|
train_dataset=tokenized_dataset,
|
|
eval_dataset=tokenized_dataset, # Używamy tego samego zbioru do ewaluacji
|
|
data_collator=data_collator
|
|
)
|
|
|
|
print("Rozpoczęcie treningu...")
|
|
trainer.train()
|
|
trainer.save_model("./trained_model/gpt")
|
|
tokenizer.save_pretrained("./trained_model/gpt")
|
|
|
|
if __name__ == "__main__":
|
|
main()
|