Ten kod działa!
This commit is contained in:
l.gabrysiak
parent
537e191d5f
commit
a0aab164cb
423
hft.py
423
hft.py
|
|
@ -1,296 +1,261 @@
|
||||||
import os
|
import os
|
||||||
import torch
|
import torch
|
||||||
import torch.nn as nn
|
import torch.nn as nn
|
||||||
|
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer, DataCollatorForLanguageModeling
|
||||||
|
from datasets import Dataset
|
||||||
import re
|
import re
|
||||||
import json
|
import json
|
||||||
import numpy as np
|
|
||||||
import PyPDF2
|
import PyPDF2
|
||||||
import docx2txt
|
import docx2txt
|
||||||
import pytesseract
|
import pytesseract
|
||||||
from PIL import Image
|
from PIL import Image
|
||||||
from collections import defaultdict
|
from collections import defaultdict
|
||||||
from transformers import (
|
|
||||||
AutoTokenizer,
|
|
||||||
AutoModelForCausalLM,
|
|
||||||
TrainingArguments,
|
|
||||||
Trainer,
|
|
||||||
DataCollatorForLanguageModeling
|
|
||||||
)
|
|
||||||
from datasets import Dataset, Features, Value
|
|
||||||
from huggingface_hub import login
|
from huggingface_hub import login
|
||||||
|
|
||||||
|
# Konfiguracja
|
||||||
|
os.environ['TORCH_USE_CUDA_DSA'] = '1'
|
||||||
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
||||||
login(token="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")
|
login(token="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")
|
||||||
|
|
||||||
class LegalAITrainer:
|
class SourceMapper:
|
||||||
def __init__(self):
|
def __init__(self):
|
||||||
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
self.source_to_idx = defaultdict(lambda: len(self.source_to_idx))
|
||||||
|
self.idx_to_source = {}
|
||||||
|
|
||||||
class SourceMapper:
|
def add_source(self, source):
|
||||||
def __init__(self):
|
if source and source not in self.source_to_idx:
|
||||||
self.source_to_idx = defaultdict(lambda: len(self.source_to_idx))
|
idx = self.source_to_idx[source]
|
||||||
self.idx_to_source = {}
|
self.idx_to_source[idx] = source
|
||||||
|
|
||||||
def add_source(self, source):
|
def get_idx(self, source):
|
||||||
if source and source not in self.source_to_idx:
|
return self.source_to_idx[source] if source else -1
|
||||||
idx = self.source_to_idx[source]
|
|
||||||
self.idx_to_source[idx] = source
|
|
||||||
|
|
||||||
def get_idx(self, source):
|
def get_source(self, idx):
|
||||||
return self.source_to_idx[source] if source else -1
|
return self.idx_to_source.get(idx, "Unknown")
|
||||||
|
|
||||||
def get_source(self, idx):
|
def load_file_catalog(catalog_path):
|
||||||
return self.idx_to_source.get(idx, "Unknown")
|
try:
|
||||||
|
with open(catalog_path, 'r', encoding='utf-8') as file:
|
||||||
|
return json.load(file)
|
||||||
|
except Exception as e:
|
||||||
|
print(f"Błąd wczytywania katalogu plików: {str(e)}")
|
||||||
|
return {}
|
||||||
|
|
||||||
class LegalModel(nn.Module):
|
def identify_legal_document(filename, file_catalog):
|
||||||
def __init__(self, model_name, config):
|
base_name = os.path.splitext(filename)[0].lower()
|
||||||
super().__init__()
|
return file_catalog.get(base_name, "Opracowanie własne")
|
||||||
self.base_model = AutoModelForCausalLM.from_pretrained(model_name, config=config)
|
|
||||||
self.source_embedding = nn.Embedding(100000, config.hidden_size, padding_idx=-1)
|
|
||||||
self.confidence_layer = nn.Linear(config.hidden_size, 1)
|
|
||||||
|
|
||||||
for param in self.base_model.parameters():
|
def extract_text_from_file(file_path):
|
||||||
param.requires_grad = False
|
try:
|
||||||
|
_, ext = os.path.splitext(file_path)
|
||||||
|
ext = ext.lower()
|
||||||
|
|
||||||
for layer in [self.source_embedding, self.confidence_layer]:
|
if ext in ['.txt', '.md']:
|
||||||
for param in layer.parameters():
|
with open(file_path, 'r', encoding='utf-8') as file:
|
||||||
param.requires_grad = True
|
return file.read()
|
||||||
|
elif ext == '.pdf':
|
||||||
def forward(self, input_ids=None, attention_mask=None, labels=None, source_idx=None):
|
text = ""
|
||||||
if source_idx is not None:
|
try:
|
||||||
source_idx = torch.clamp(source_idx, 0, self.source_embedding.num_embeddings-1)
|
with open(file_path, 'rb') as file:
|
||||||
source_embeds = self.source_embedding(source_idx).unsqueeze(1)
|
reader = PyPDF2.PdfReader(file)
|
||||||
inputs_embeds = self.base_model.get_input_embeddings()(input_ids) + source_embeds
|
|
||||||
outputs = self.base_model(
|
|
||||||
inputs_embeds=inputs_embeds,
|
|
||||||
attention_mask=attention_mask,
|
|
||||||
labels=labels
|
|
||||||
)
|
|
||||||
else:
|
|
||||||
outputs = self.base_model(
|
|
||||||
input_ids=input_ids,
|
|
||||||
attention_mask=attention_mask,
|
|
||||||
labels=labels
|
|
||||||
)
|
|
||||||
|
|
||||||
confidence = torch.sigmoid(self.confidence_layer(outputs.hidden_states[-1].mean(dim=1)))
|
|
||||||
return {
|
|
||||||
"loss": outputs.loss,
|
|
||||||
"logits": outputs.logits,
|
|
||||||
"confidence": confidence,
|
|
||||||
"hidden_states": outputs.hidden_states
|
|
||||||
}
|
|
||||||
|
|
||||||
def load_file_catalog(self, catalog_path):
|
|
||||||
try:
|
|
||||||
with open(catalog_path, 'r', encoding='utf-8') as f:
|
|
||||||
return json.load(f)
|
|
||||||
except Exception as e:
|
|
||||||
print(f"Błąd ładowania katalogu: {str(e)}")
|
|
||||||
return {}
|
|
||||||
|
|
||||||
def extract_text(self, file_path):
|
|
||||||
ext = os.path.splitext(file_path)[1].lower()
|
|
||||||
try:
|
|
||||||
if ext in ['.txt', '.md']:
|
|
||||||
with open(file_path, 'r', encoding='utf-8') as f:
|
|
||||||
return f.read()
|
|
||||||
elif ext == '.pdf':
|
|
||||||
text = ""
|
|
||||||
with open(file_path, 'rb') as f:
|
|
||||||
reader = PyPDF2.PdfReader(f)
|
|
||||||
for page in reader.pages:
|
for page in reader.pages:
|
||||||
text += page.extract_text() or ""
|
text += page.extract_text() or ""
|
||||||
return text
|
except Exception as e:
|
||||||
elif ext in ['.doc', '.docx']:
|
print(f"Błąd PDF: {str(e)}")
|
||||||
return docx2txt.process(file_path)
|
return text
|
||||||
elif ext in ['.jpg', '.jpeg', '.png']:
|
elif ext in ['.doc', '.docx']:
|
||||||
return pytesseract.image_to_string(Image.open(file_path))
|
return docx2txt.process(file_path)
|
||||||
else:
|
elif ext in ['.jpg', '.jpeg', '.png', '.bmp', '.tiff']:
|
||||||
return ""
|
return pytesseract.image_to_string(Image.open(file_path))
|
||||||
except Exception as e:
|
else:
|
||||||
print(f"Błąd przetwarzania {file_path}: {str(e)}")
|
print(f"Nieobsługiwany format pliku: {ext}")
|
||||||
return ""
|
return ""
|
||||||
|
except Exception as e:
|
||||||
|
print(f"Błąd ekstrakcji tekstu: {str(e)}")
|
||||||
|
return ""
|
||||||
|
|
||||||
def prepare_data(self, data_dir, catalog_path):
|
def prepare_dataset(directory, catalog_path, source_mapper):
|
||||||
catalog = self.load_file_catalog(catalog_path)
|
file_catalog = load_file_catalog(catalog_path)
|
||||||
data = []
|
data = []
|
||||||
source_mapper = self.SourceMapper()
|
|
||||||
|
|
||||||
for root, _, files in os.walk(data_dir):
|
print(f"\n{'='*50}\nDIAGNOSTYKA DANYCH\n{'='*50}")
|
||||||
for file in files:
|
|
||||||
file_path = os.path.join(root, file)
|
|
||||||
text = self.extract_text(file_path)
|
|
||||||
|
|
||||||
if not text:
|
for root, _, files in os.walk(directory):
|
||||||
|
for file in files:
|
||||||
|
file_path = os.path.join(root, file)
|
||||||
|
print(f"\nPrzetwarzanie pliku: {file_path}")
|
||||||
|
|
||||||
|
try:
|
||||||
|
text = extract_text_from_file(file_path)
|
||||||
|
if not text.strip():
|
||||||
|
print("Pominięto - brak tekstu")
|
||||||
continue
|
continue
|
||||||
|
|
||||||
doc_type = catalog.get(os.path.splitext(file)[0].lower(), "Opracowanie własne")
|
print(f"Długość tekstu: {len(text)} znaków")
|
||||||
|
|
||||||
|
doc_type = identify_legal_document(file, file_catalog)
|
||||||
|
print(f"Rozpoznany typ dokumentu: {doc_type}")
|
||||||
|
|
||||||
if doc_type != "Opracowanie własne":
|
if doc_type != "Opracowanie własne":
|
||||||
articles = re.split(r'(?i)(Art\.\s*\d+[a-z]*)', text)
|
articles = re.split(r'(?i)(Art[\.\s]+\d+[\.\s]?)', text)
|
||||||
for i in range(1, len(articles), 2):
|
articles = [a.strip() for a in articles if a.strip()]
|
||||||
art_num = articles[i].strip()
|
|
||||||
content = articles[i+1].strip()
|
|
||||||
|
|
||||||
if len(content) < 100:
|
print(f"Znaleziono {len(articles)} fragmentów")
|
||||||
|
|
||||||
|
for i in range(0, len(articles)-1, 2):
|
||||||
|
article_number = articles[i]
|
||||||
|
article_content = articles[i+1]
|
||||||
|
|
||||||
|
if len(article_content) < 50:
|
||||||
continue
|
continue
|
||||||
|
|
||||||
source = f"{doc_type}, {art_num}"
|
source = f"{doc_type}, {article_number}"
|
||||||
source_mapper.add_source(source)
|
source_mapper.add_source(source)
|
||||||
data.append({
|
data.append({
|
||||||
"text": f"[LEGAL] {art_num} {content}",
|
"text": f"{article_number} {article_content}",
|
||||||
"source_idx": source_mapper.get_idx(source),
|
"source_idx": source_mapper.get_idx(source)
|
||||||
"is_legal": 1
|
|
||||||
})
|
})
|
||||||
else:
|
else:
|
||||||
chunks = [f"[GENERAL] {text[i:i+512]}" for i in range(0, len(text), 512)]
|
clean_text = re.sub(r'\s+', ' ', text).strip()
|
||||||
|
chunks = [clean_text[i:i+512] for i in range(0, len(clean_text), 512)]
|
||||||
|
chunks = [c for c in chunks if c.strip()]
|
||||||
|
|
||||||
for chunk in chunks:
|
for chunk in chunks:
|
||||||
data.append({
|
data.append({
|
||||||
"text": chunk,
|
"text": chunk,
|
||||||
"source_idx": -1,
|
"source_idx": -1
|
||||||
"is_legal": 0
|
|
||||||
})
|
})
|
||||||
|
print(f"Dodano {len(chunks)} chunków")
|
||||||
|
|
||||||
features = Features({
|
except Exception as e:
|
||||||
"text": Value("string"),
|
print(f"Błąd podczas przetwarzania pliku: {str(e)}")
|
||||||
"source_idx": Value("int32"),
|
continue
|
||||||
"is_legal": Value("int32")
|
|
||||||
})
|
|
||||||
|
|
||||||
return Dataset.from_dict({
|
print(f"\nPodsumowanie przygotowania danych:")
|
||||||
"text": [d["text"] for d in data],
|
print(f"Łączna liczba przykładów: {len(data)}")
|
||||||
"source_idx": np.array([d["source_idx"] for d in data], dtype=np.int32),
|
if data:
|
||||||
"is_legal": np.array([d["is_legal"] for d in data], dtype=np.int32)
|
print("Przykładowy wpis:")
|
||||||
}, features=features), source_mapper
|
print(json.dumps(data[0], indent=2, ensure_ascii=False))
|
||||||
|
else:
|
||||||
|
print("BRAK DANYCH - sprawdź diagnostykę powyżej")
|
||||||
|
|
||||||
def train(self, model_name="crumb/nano-mistral", data_dir="data", catalog_path="catalog.json"):
|
return data
|
||||||
dataset, source_mapper = self.prepare_data(data_dir, catalog_path)
|
|
||||||
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
|
||||||
tokenizer.pad_token = tokenizer.eos_token
|
|
||||||
|
|
||||||
def tokenize_fn(examples):
|
class CustomModel(nn.Module):
|
||||||
tokenized = tokenizer(
|
def __init__(self, model_name, config):
|
||||||
examples["text"],
|
super().__init__()
|
||||||
padding="max_length",
|
self.base_model = AutoModelForCausalLM.from_pretrained(model_name, config=config)
|
||||||
truncation=True,
|
self.source_embedding = nn.Embedding(10000, config.hidden_size, padding_idx=-1)
|
||||||
max_length=512,
|
|
||||||
return_tensors="pt"
|
for param in self.base_model.parameters():
|
||||||
|
param.requires_grad = False
|
||||||
|
for param in self.base_model.get_output_embeddings().parameters():
|
||||||
|
param.requires_grad = True
|
||||||
|
|
||||||
|
def forward(self, input_ids=None, attention_mask=None, labels=None, source_idx=None, **kwargs):
|
||||||
|
if source_idx is not None:
|
||||||
|
valid_indices = torch.clamp(source_idx, 0, self.source_embedding.num_embeddings-1)
|
||||||
|
source_embeds = self.source_embedding(valid_indices).unsqueeze(1)
|
||||||
|
inputs_embeds = self.base_model.get_input_embeddings()(input_ids) + source_embeds
|
||||||
|
return self.base_model(
|
||||||
|
inputs_embeds=inputs_embeds,
|
||||||
|
attention_mask=attention_mask,
|
||||||
|
labels=labels,
|
||||||
|
**kwargs
|
||||||
)
|
)
|
||||||
return {
|
return self.base_model(
|
||||||
"input_ids": tokenized["input_ids"].squeeze().tolist(),
|
input_ids=input_ids,
|
||||||
"attention_mask": tokenized["attention_mask"].squeeze().tolist(),
|
attention_mask=attention_mask,
|
||||||
"labels": tokenized["input_ids"].squeeze().clone().tolist(),
|
labels=labels,
|
||||||
"source_idx": examples["source_idx"]
|
**kwargs
|
||||||
}
|
|
||||||
|
|
||||||
tokenized_dataset = dataset.map(tokenize_fn, batched=True, batch_size=16)
|
|
||||||
|
|
||||||
class CustomDataCollator(DataCollatorForLanguageModeling):
|
|
||||||
def torch_call(self, examples):
|
|
||||||
batch = super().torch_call(examples)
|
|
||||||
if "source_idx" in examples[0]:
|
|
||||||
batch["source_idx"] = torch.tensor(
|
|
||||||
[ex["source_idx"] for ex in examples],
|
|
||||||
dtype=torch.int32
|
|
||||||
)
|
|
||||||
return batch
|
|
||||||
|
|
||||||
config = AutoModelForCausalLM.from_pretrained(model_name).config
|
|
||||||
model = self.LegalModel(model_name, config).to(self.device)
|
|
||||||
|
|
||||||
training_args = TrainingArguments(
|
|
||||||
output_dir="./legal_ai_model",
|
|
||||||
num_train_epochs=3,
|
|
||||||
per_device_train_batch_size=2,
|
|
||||||
gradient_accumulation_steps=4,
|
|
||||||
learning_rate=2e-5,
|
|
||||||
fp16=torch.cuda.is_available(),
|
|
||||||
logging_steps=50,
|
|
||||||
save_strategy="steps",
|
|
||||||
save_steps=500,
|
|
||||||
report_to="none",
|
|
||||||
remove_unused_columns=False
|
|
||||||
)
|
)
|
||||||
|
|
||||||
class LegalTrainer(Trainer):
|
def generate(self, *args, **kwargs):
|
||||||
def compute_loss(self, model, inputs, return_outputs=False):
|
return self.base_model.generate(*args, **kwargs)
|
||||||
outputs = model(**inputs)
|
|
||||||
loss = outputs["loss"]
|
|
||||||
|
|
||||||
target_conf = (inputs["source_idx"] != -1).float()
|
class CustomDataCollator(DataCollatorForLanguageModeling):
|
||||||
conf_loss = nn.BCELoss()(outputs["confidence"].squeeze(), target_conf)
|
def torch_call(self, examples):
|
||||||
|
# Przetwórz podstawowe pola
|
||||||
|
input_ids = torch.stack([torch.tensor(ex["input_ids"]) for ex in examples])
|
||||||
|
attention_mask = torch.stack([torch.tensor(ex["attention_mask"]) for ex in examples])
|
||||||
|
labels = torch.stack([torch.tensor(ex["labels"]) for ex in examples])
|
||||||
|
|
||||||
total_loss = loss + 0.7 * conf_loss
|
batch = {
|
||||||
return (total_loss, outputs) if return_outputs else total_loss
|
"input_ids": input_ids,
|
||||||
|
"attention_mask": attention_mask,
|
||||||
|
"labels": labels
|
||||||
|
}
|
||||||
|
|
||||||
trainer = LegalTrainer(
|
# Dodaj source_idx jeśli istnieje
|
||||||
model=model,
|
if "source_idx" in examples[0]:
|
||||||
args=training_args,
|
source_idx = torch.stack([torch.tensor(ex["source_idx"]) for ex in examples])
|
||||||
train_dataset=tokenized_dataset,
|
batch["source_idx"] = source_idx
|
||||||
data_collator=CustomDataCollator(tokenizer=tokenizer, mlm=False)
|
|
||||||
)
|
|
||||||
|
|
||||||
print("Rozpoczęcie treningu...")
|
return batch
|
||||||
trainer.train()
|
|
||||||
|
|
||||||
model.save_pretrained("./trained_legal_ai")
|
def main():
|
||||||
tokenizer.save_pretrained("./trained_legal_ai")
|
source_mapper = SourceMapper()
|
||||||
with open("./trained_legal_ai/source_mapper.json", "w") as f:
|
model_name = "crumb/nano-mistral"
|
||||||
json.dump(source_mapper.idx_to_source, f)
|
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||||||
|
tokenizer.pad_token = tokenizer.eos_token
|
||||||
|
|
||||||
print("Trening zakończony!")
|
# Przygotowanie danych
|
||||||
|
catalog_path = "file_catalog.json"
|
||||||
|
data = prepare_dataset("files", catalog_path, source_mapper)
|
||||||
|
|
||||||
def generate_response(self, prompt, confidence_threshold=0.65):
|
if not data:
|
||||||
model = self.LegalModel.from_pretrained(
|
print("\nBrak danych do treningu!")
|
||||||
"./trained_legal_ai",
|
return
|
||||||
config=AutoModelForCausalLM.from_pretrained("crumb/nano-mistral").config
|
|
||||||
).to(self.device)
|
|
||||||
|
|
||||||
tokenizer = AutoTokenizer.from_pretrained("./trained_legal_ai")
|
#dataset = Dataset.from_list(data)
|
||||||
|
dataset = Dataset.from_dict({k: [d[k] for d in data] for k in data[0]})
|
||||||
|
|
||||||
with open("./trained_legal_ai/source_mapper.json", "r") as f:
|
|
||||||
source_mapper = json.load(f)
|
|
||||||
|
|
||||||
inputs = tokenizer(
|
def tokenize_function(examples):
|
||||||
f"[PROMPT] {prompt} [RESPONSE]",
|
tokenized = tokenizer(
|
||||||
return_tensors="pt",
|
examples["text"],
|
||||||
|
truncation=True,
|
||||||
|
padding="max_length",
|
||||||
max_length=512,
|
max_length=512,
|
||||||
truncation=True
|
return_tensors="pt"
|
||||||
).to(self.device)
|
)
|
||||||
|
return {
|
||||||
|
"input_ids": tokenized["input_ids"].squeeze(),
|
||||||
|
"attention_mask": tokenized["attention_mask"].squeeze(),
|
||||||
|
"labels": tokenized["input_ids"].squeeze().clone(),
|
||||||
|
"source_idx": examples["source_idx"] # Dodano bez konwersji do tensora
|
||||||
|
}
|
||||||
|
|
||||||
with torch.no_grad():
|
tokenized_dataset = dataset.map(tokenize_function, batched=True, batch_size=16)
|
||||||
outputs = model.generate(
|
|
||||||
input_ids=inputs.input_ids,
|
|
||||||
attention_mask=inputs.attention_mask,
|
|
||||||
max_length=512,
|
|
||||||
do_sample=True,
|
|
||||||
temperature=0.7,
|
|
||||||
top_k=50,
|
|
||||||
pad_token_id=tokenizer.eos_token_id,
|
|
||||||
output_scores=True,
|
|
||||||
return_dict_in_generate=True
|
|
||||||
)
|
|
||||||
|
|
||||||
full_text = tokenizer.decode(outputs.sequences[0], skip_special_tokens=True)
|
model = CustomModel(model_name, AutoModelForCausalLM.from_pretrained(model_name).config)
|
||||||
confidence = torch.sigmoid(outputs.scores[-1][:, tokenizer.eos_token_id]).item()
|
model.source_mapper = source_mapper
|
||||||
|
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
||||||
|
model.to(device)
|
||||||
|
|
||||||
citations = list(set(re.findall(r"Art\.\s*\d+[a-z]*", full_text)))
|
training_args = TrainingArguments(
|
||||||
verified = [c for c in citations if any(c in s for s in source_mapper.values())]
|
output_dir="./results",
|
||||||
|
num_train_epochs=3,
|
||||||
if confidence < confidence_threshold or not verified:
|
per_device_train_batch_size=2,
|
||||||
return "Nie mogę udzielić jednoznacznej odpowiedzi na podstawie dostępnych danych."
|
gradient_accumulation_steps=4,
|
||||||
else:
|
learning_rate=2e-5,
|
||||||
return f"{full_text}\n\nPotwierdzone źródła: {', '.join(verified)}"
|
fp16=torch.cuda.is_available(),
|
||||||
|
logging_steps=10,
|
||||||
if __name__ == "__main__":
|
save_strategy="steps",
|
||||||
legal_ai = LegalAITrainer()
|
save_steps=1000,
|
||||||
|
report_to="none",
|
||||||
legal_ai.train(
|
remove_unused_columns=False
|
||||||
model_name="crumb/nano-mistral",
|
|
||||||
data_dir="./legal_docs",
|
|
||||||
catalog_path="./catalog.json"
|
|
||||||
)
|
)
|
||||||
|
|
||||||
test_prompt = "Jakie są kary za nieprzestrzeganie przepisów RODO?"
|
trainer = Trainer(
|
||||||
print(legal_ai.generate_response(test_prompt))
|
model=model,
|
||||||
|
args=training_args,
|
||||||
|
train_dataset=tokenized_dataset,
|
||||||
|
data_collator=CustomDataCollator(tokenizer=tokenizer, mlm=False)
|
||||||
|
)
|
||||||
|
|
||||||
|
print("\nRozpoczęcie treningu...")
|
||||||
|
trainer.train()
|
||||||
|
|
||||||
|
if __name__ == "__main__":
|
||||||
|
main()
|
||||||
Loading…
Reference in New Issue