mod allegro

allegro.py

@@ -1,12 +1,17 @@
 import os
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
 import torch
-import weaviate
+import faiss
-import numpy as np
-from datasets import Dataset
-from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, TrainingArguments, Trainer, DataCollatorForSeq2Seq
 from weaviate.connect import ConnectionParams
 import weaviate
-import tempfile
+
+from sentence_transformers import SentenceTransformer
+from datasets import Dataset
+from peft import LoraConfig, get_peft_model
+from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer, DataCollatorForLanguageModeling
+
+embed_model = SentenceTransformer("all-MiniLM-L6-v2")
 
 # 1️⃣ Połączenie z Weaviate
 client = weaviate.WeaviateClient(
@@ -38,67 +43,80 @@ def fetch_documents():
 
 documents = fetch_documents()
 
-# 3️⃣ Inicjalizacja modelu
+embeddings = embed_model.encode(documents)
-model_name = "allegro/multislav-5lang"
-device = "cuda" if torch.cuda.is_available() else "cpu"
 
-model = AutoModelForSeq2SeqLM.from_pretrained(model_name).to(device)
+dim = embeddings.shape[1]
-tokenizer = AutoTokenizer.from_pretrained(model_name)
+index = faiss.IndexFlatL2(dim)
+index.add(np.array(embeddings, dtype=np.float32))
 
-# 4️⃣ Przygotowanie danych treningowych
 def create_training_data():
-    return Dataset.from_dict({"text": documents})
+    data = {
+        "text": documents,
+        "embedding": embeddings.tolist()
+    }
+    return Dataset.from_dict(data)
 
 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
+device = "cuda" if torch.cuda.is_available() else "cpu"
+model_name = "allegro/multislav-5lang"
+model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16).to(device)
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+lora_config = LoraConfig(
+    r=8, lora_alpha=32, lora_dropout=0.1, bias="none", task_type="CAUSAL_LM"
+)
+model = get_peft_model(model, lora_config)
+
+max_length = 384
+
 def tokenize_function(examples):
     return tokenizer(
         examples["text"],
         padding="max_length",
         truncation=True,
-        max_length=512
+        max_length=max_length
     )
 
 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_strategy="steps",          # Ewaluacja co określoną liczbę kroków
-    eval_steps=500,
+    eval_steps=500,                 # Ewaluacja co 500 kroków
-    save_steps=500,
+    save_strategy="steps",          # Zapis modelu co określoną liczbę kroków
-    learning_rate=2e-5,
+    save_steps=500,                 # Zapis modelu co 500 kroków
+    learning_rate=1e-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,    # Wczytaj najlepszy model na końcu
-    load_best_model_at_end=True,
+    metric_for_best_model="loss",   # Kryterium wyboru najlepszego modelu
-    metric_for_best_model="loss",
+    greater_is_better=False,        # Niższy loss = lepszy model
-    greater_is_better=False,
 )
 
-# 7️⃣ Data Collator
+data_collator = DataCollatorForLanguageModeling(
-data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, model=model)
+    tokenizer=tokenizer,
+    mlm=False
+)
 
-# 8️⃣ Trening
 trainer = Trainer(
     model=model,
     args=training_args,
     train_dataset=tokenized_train,
-    eval_dataset=tokenized_eval,
+    eval_dataset=tokenized_eval,    # Dodany zestaw ewaluacyjny
     data_collator=data_collator,
 )
 
 trainer.train()
 
-# 9️⃣ Zapis modelu
+model.save_pretrained("./trained_model/gemma")
-model.save_pretrained("./models/allegro")
+tokenizer.save_pretrained("./trained_model/gemma")
-tokenizer.save_pretrained("./models/allegro")
 
 print("✅ Model został wytrenowany i zapisany!")