mod herbert

herbert.py

@@ -6,21 +6,28 @@ import faiss
 import numpy as np
 from sentence_transformers import SentenceTransformer
 from datasets import Dataset
-from peft import LoraConfig, get_peft_model, PeftModel
-from transformers import (AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer, 
-                          DataCollatorForLanguageModeling, LlamaTokenizer, LlamaForCausalLM)
-import bitsandbytes as bnb
+from peft import LoraConfig, get_peft_model
+from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer, DataCollatorForLanguageModeling
 
 # 1️⃣ Inicjalizacja modelu do embeddingów
 embed_model = SentenceTransformer("all-MiniLM-L6-v2")
 
-# 2️⃣ Wczytanie dokumentów i embeddingów
+# 2️⃣ Dodanie dokumentów i embeddingów
 def read_documents_from_file(file_path):
     with open(file_path, 'r', encoding='utf-8') as file:
         content = file.read()
         articles = content.split('\n\n')
-        return [article.strip() for article in articles if article.strip().startswith('Art.')]
-
+        documents = []
+        for article in articles:
+            if article.strip().startswith('Art.'):
+                documents.append(article.strip())
+    return documents
+#documents = [
+#    "Jak założyć firmę w Polsce?", 
+#    "Jak rozliczyć podatek VAT?", 
+#    "Procedura składania reklamacji w e-sklepie.",
+#    "Jakie dokumenty są potrzebne do rejestracji działalności?"
+#]
 file_path = './docs/kodekspracy.txt'  # Zmień na właściwą ścieżkę
 documents = read_documents_from_file(file_path)
 embeddings = embed_model.encode(documents)
@@ -32,31 +39,41 @@ index.add(np.array(embeddings, dtype=np.float32))
 
 # 4️⃣ Przygotowanie danych treningowych
 def create_training_data():
-    return Dataset.from_dict({"text": documents, "embedding": embeddings.tolist()})
+    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, eval_dataset = split_dataset["train"], split_dataset["test"]
+train_dataset = split_dataset["train"]
+eval_dataset = split_dataset["test"]
 
-# 5️⃣ Ładowanie modelu bazowego i fine-tunowanego
-base_model = "decapoda-research/llama-7b-hf"
-finetuned_model = "mmosiolek/polpaca-lora-7b"
-
-tokenizer = LlamaTokenizer.from_pretrained(base_model)
-tokenizer.pad_token_id = 0
-tokenizer.padding_side = "left"
-
-model = LlamaForCausalLM.from_pretrained(base_model, torch_dtype=torch.float16).to("cuda")
-model = PeftModel.from_pretrained(model, finetuned_model).to("cuda")
+# 5️⃣ Ładowanie modelu Gemma 2B
+device = "cuda" if torch.cuda.is_available() else "cpu"
+model_name = "Lajonbot/vicuna-7b-v1.5-PL-lora_unload"
+model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16).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="CAUSAL_LM")
+    r=8, lora_alpha=32, lora_dropout=0.1, bias="none", task_type="CAUSAL_LM"
+)
 model = get_peft_model(model, lora_config)
 
-# 7️⃣ Tokenizacja
+# 7️⃣ Tokenizacja danych
+max_length = 384
+
 def tokenize_function(examples):
-    return tokenizer(examples["text"], padding="max_length", truncation=True, max_length=384)
+    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)
@@ -64,35 +81,39 @@ tokenized_eval = eval_dataset.map(tokenize_function, batched=True)
 # 8️⃣ Parametry treningu
 training_args = TrainingArguments(
     output_dir="./results",
-    evaluation_strategy="steps",
-    eval_steps=500,
-    save_strategy="steps",
-    save_steps=500,
+    eval_strategy="steps",          # Ewaluacja co określoną liczbę kroków
+    eval_steps=500,                 # Ewaluacja co 500 kroków
+    save_strategy="steps",          # Zapis modelu co określoną liczbę kroków
+    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,
-    load_best_model_at_end=True,
-    metric_for_best_model="loss",
-    greater_is_better=False,
+    load_best_model_at_end=True,    # Wczytaj najlepszy model na końcu
+    metric_for_best_model="loss",   # Kryterium wyboru najlepszego modelu
+    greater_is_better=False,        # Niższy loss = lepszy model
 )
 
 # 9️⃣ Data Collator
-data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
+data_collator = DataCollatorForLanguageModeling(
+    tokenizer=tokenizer,
+    mlm=False
+)
 
-# 🔟 Trening
+# 🔟 Trening modelu
 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()
 
-# 1️⃣1️⃣ Zapis modelu lokalnie
-model.save_pretrained("./models/finetuned_llama")
-tokenizer.save_pretrained("./models/finetuned_llama")
+# 1️⃣1️⃣ Zapis modelu
+model.save_pretrained("./models/herbert")
+tokenizer.save_pretrained("./models/herbert")
 
 print("✅ Model został wytrenowany i zapisany!")