ably.do/allegro.py

import os
import torch
import weaviate
import numpy as np
from datasets import Dataset
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, TrainingArguments, Trainer, DataCollatorForSeq2Seq
from weaviate.connect import ConnectionParams
import weaviate

# 1️⃣ Połączenie z Weaviate
client = weaviate.WeaviateClient(
    connection_params=ConnectionParams.from_params(
        http_host="weaviate",
        http_port=8080,
        http_secure=False,
        grpc_host="weaviate",
        grpc_port=50051,
        grpc_secure=False,
    )
)

# 2️⃣ Pobranie dokumentów z Weaviate
def fetch_documents():
    query = client.query.get("Document", ["content", "fileName"]).do()
    documents = []
    for item in query["data"]["Get"]["Document"]:
        file_name = item.get("fileName", "unknown_file")
        content = item.get("content", "")
        if content:
            documents.append(f"fileName: {file_name}, content: {content}")
    return documents

documents = fetch_documents()

# 3️⃣ Inicjalizacja modelu
model_name = "allegro/multislav-5lang"
device = "cuda" if torch.cuda.is_available() else "cpu"

model = AutoModelForSeq2SeqLM.from_pretrained(model_name).to(device)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# 4️⃣ Przygotowanie danych treningowych
def create_training_data():
    return Dataset.from_dict({"text": documents})

dataset = create_training_data()
split_dataset = dataset.train_test_split(test_size=0.25)
train_dataset = split_dataset["train"]
eval_dataset = split_dataset["test"]

# 5️⃣ Tokenizacja
def tokenize_function(examples):
    return tokenizer(
        examples["text"],
        padding="max_length",
        truncation=True,
        max_length=512
    )

tokenized_train = train_dataset.map(tokenize_function, batched=True)
tokenized_eval = eval_dataset.map(tokenize_function, batched=True)

# 6️⃣ Parametry treningu
training_args = TrainingArguments(
    output_dir="./results",
    evaluation_strategy="steps",
    eval_steps=500,
    save_steps=500,
    learning_rate=2e-5,
    per_device_train_batch_size=2,
    per_device_eval_batch_size=2,
    num_train_epochs=16,
    weight_decay=0.01,
    save_total_limit=2,
    load_best_model_at_end=True,
    metric_for_best_model="loss",
    greater_is_better=False,
)

# 7️⃣ Data Collator
data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, model=model)

# 8️⃣ Trening
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_train,
    eval_dataset=tokenized_eval,
    data_collator=data_collator,
)

trainer.train()

# 9️⃣ Zapis modelu
model.save_pretrained("./models/allegro")
tokenizer.save_pretrained("./models/allegro")

print("✅ Model został wytrenowany i zapisany!")