From 51828191cc6acc38817353479d1edff3dc80972a Mon Sep 17 00:00:00 2001
From: "l.gabrysiak" <l.gabrysiak@osadkowski.pl>
Date: Wed, 26 Feb 2025 13:17:57 +0100
Subject: [PATCH] mod gemma

---
 gemma.py | 19 ++++++++++---------
 1 file changed, 10 insertions(+), 9 deletions(-)

diff --git a/gemma.py b/gemma.py
index aa58c21..86f6bd1 100644
--- a/gemma.py
+++ b/gemma.py
@@ -64,17 +64,18 @@ tokenized_dataset = dataset.map(tokenize_function, batched=True)
 # 8️⃣ Parametry treningu
 training_args = TrainingArguments(
     output_dir="./results",
-    per_device_train_batch_size=2,
-    gradient_accumulation_steps=4,  # Symuluje większy batch size
-    num_train_epochs=5,
-    logging_dir="./logs",
-    save_strategy="epoch",
+    evaluation_strategy="steps",  # Zmienione na "steps"
+    eval_steps=500,  # Dodane
+    save_strategy="steps",  # Zmienione na "steps"
+    save_steps=500,  # Dodane, musi być takie samo jak eval_steps lub jego wielokrotność
     learning_rate=2e-5,
-    warmup_steps=100,
-    fp16=True,  # Używa mixed precision training
-    evaluation_strategy="steps",
-    eval_steps=500,
+    per_device_train_batch_size=2,
+    per_device_eval_batch_size=2,
+    num_train_epochs=5,
+    weight_decay=0.01,
     load_best_model_at_end=True,
+    metric_for_best_model="loss",  # lub inna metryka, którą chcesz optymalizować
+    greater_is_better=False,  # Ustaw na True, jeśli wyższa wartość metryki jest lepsza
 )
 
 # 9️⃣ Data Collator