mod

hft.py

@@ -15,7 +15,6 @@ from huggingface_hub import login
 login(token="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")
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
 
-# Nowa klasa do zarządzania źródłami
 class SourceMapper:
     def __init__(self):
         self.source_to_idx = defaultdict(lambda: len(self.source_to_idx))
@@ -119,7 +118,7 @@ class CustomModel(nn.Module):
         super().__init__()
         self.base_model = AutoModelForCausalLM.from_pretrained(model_name, config=config)
         self.source_embedding = nn.Embedding(
-            num_embeddings=1000,  # Maksymalna liczba unikalnych źródeł
+            num_embeddings=1000,
             embedding_dim=config.hidden_size,
             padding_idx=-1
         )
@@ -133,8 +132,9 @@ class CustomModel(nn.Module):
         )
         
         if source_idx is not None:
-            # Dodaj embedding źródła do logits
-            source_embeds = self.source_embedding(source_idx).unsqueeze(1)
+            print("outputs.logits shape:", outputs.logits.shape)
+            source_embeds = self.source_embedding(source_idx).unsqueeze(1).expand(-1, outputs.logits.size(1), -1)
+            print("source_embeds shape:", source_embeds.shape)
             outputs.logits += source_embeds
         
         return outputs
@@ -149,7 +149,7 @@ class CustomTrainer(Trainer):
 
 # Inicjalizacja komponentów
 source_mapper = SourceMapper()
-model_name = "crumb/nano-mistral" #"google/gemma-2-2b"
+model_name = "crumb/nano-mistral"
 tokenizer = AutoTokenizer.from_pretrained(model_name)
 tokenizer.pad_token = tokenizer.eos_token
 
@@ -162,6 +162,7 @@ tokenized_dataset = dataset.map(tokenize_function, batched=True, batch_size=8)
 # Inicjalizacja modelu
 config = AutoModelForCausalLM.from_pretrained(model_name).config
 model = CustomModel(model_name, config)
+model.to("cpu")
 
 # Konfiguracja treningu
 training_args = TrainingArguments(
@@ -171,11 +172,10 @@ training_args = TrainingArguments(
     gradient_accumulation_steps=4,
     learning_rate=2e-5,
     fp16=True,
-    logging_steps=1,  # Częstsze logowanie
-    logging_dir="./logs",  # Katalog na logi
+    logging_steps=1,
+    logging_dir="./logs",
     save_strategy="steps",
     save_steps=1000,
-    #report_to="none"
 )
 
 # Trening
@@ -183,7 +183,7 @@ trainer = CustomTrainer(
     model=model,
     args=training_args,
     train_dataset=tokenized_dataset,
-    data_collator=custom_collate_fn,  # Użyj niestandardowego collate_fn
+    data_collator=custom_collate_fn,
 )
 trainer.train()
 
@@ -211,4 +211,4 @@ def generate_answer(question, model, tokenizer, source_mapper, max_length=200):
 # Przykład użycia
 question = "Ile dni urlopu przysługuje pracownikowi?"
 answer = generate_answer(question, model, tokenizer, source_mapper)
-print(answer)
+print(answer)