ably.do/allegro.py

import os
os.environ["TOKENIZERS_PARALLELISM"] = "false"

import torch
import numpy as np
from sentence_transformers import SentenceTransformer
from datasets import Dataset
from peft import LoraConfig, get_peft_model
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, TrainingArguments, Trainer, DataCollatorForSeq2Seq
import weaviate
from weaviate.client import WeaviateClient
from weaviate.connect import ConnectionParams

# 1️⃣ Inicjalizacja modelu do embeddingów
embed_model = SentenceTransformer("all-MiniLM-L6-v2")

# 2️⃣ Połączenie z Weaviate i pobranie dokumentów
client = WeaviateClient(
    connection_params=ConnectionParams.from_params(
        http_host="weaviate",
        http_port=8080,
        http_secure=False,
        grpc_host="weaviate",
        grpc_port=50051,
        grpc_secure=False,
    )
)

collection_name = "Document"  # Zakładam, że to jest nazwa Twojej kolekcji
result = (
    client.query.get(collection_name, ["content"])
    .with_additional(["id"])
    .do()
)

documents = [item['content'] for item in result['data']['Get'][collection_name]]

# 3️⃣ Generowanie embeddingów
embeddings = embed_model.encode(documents)

# 4️⃣ Przygotowanie danych treningowych
def create_training_data():
    data = {
        "text": documents,
        "embedding": embeddings.tolist()
    }
    return Dataset.from_dict(data)

dataset = create_training_data()

# Podział danych na treningowe i ewaluacyjne
split_dataset = dataset.train_test_split(test_size=0.25)
train_dataset = split_dataset["train"]
eval_dataset = split_dataset["test"]

# 5️⃣ Ładowanie modelu allegro/multislav-5lang
device = "cuda" if torch.cuda.is_available() else "cpu"
model_name = "allegro/multislav-5lang"
model = AutoModelForSeq2SeqLM.from_pretrained(model_name).to(device)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# 6️⃣ Konfiguracja LoRA
lora_config = LoraConfig(
    r=8, lora_alpha=32, lora_dropout=0.1, bias="none", task_type="SEQ_2_SEQ_LM"
)
model = get_peft_model(model, lora_config)

# 7️⃣ Tokenizacja danych
max_length = 384

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

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

# 8️⃣ Parametry treningu
training_args = TrainingArguments(
    output_dir="./results",
    eval_strategy="steps",
    eval_steps=500,
    save_strategy="steps",
    save_steps=500,
    learning_rate=1e-5,
    per_device_train_batch_size=2,
    per_device_eval_batch_size=2,
    num_train_epochs=16,
    weight_decay=0.01,
    load_best_model_at_end=True,
    metric_for_best_model="loss",
    greater_is_better=False,
)

# 9️⃣ Data Collator
data_collator = DataCollatorForSeq2Seq(
    tokenizer=tokenizer,
    model=model
)

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

trainer.train()

# 1️⃣1️⃣ Zapis modelu
model.save_pretrained("./models/allegro")
tokenizer.save_pretrained("./models/allegro")

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