Upload run_sft_job.py with huggingface_hub

run_sft_job.py

@@ -39,12 +39,26 @@ from trl import SFTTrainer, SFTConfig
 
 
 
+def format_chat_template(example):
+    # This function is used to create a single string from the chat messages.
+    # It mimics the format used by the chat templates in the TRL library.
+    text = ""
+    for message in example["messages"]:
+        role = message["role"]
+        content = message["content"]
+        text += f"**{role.capitalize()}:** {content}\n\n"
+    return {"text": text}
+
 # 1. Load Dataset
 print("📦 Loading dataset OliverSlivka/itemsety-real-training...")
 dataset = load_dataset("OliverSlivka/itemsety-real-training")
+
+# Apply the formatting function to create the 'text' column
+dataset = dataset.map(format_chat_template)
+
 train_dataset = dataset["train"]
 eval_dataset = dataset["validation"]
-print(f"✅ Dataset loaded. Train: {len(train_dataset)}, Eval: {len(eval_dataset)}")
+print(f"✅ Dataset loaded and formatted. Train: {len(train_dataset)}, Eval: {len(eval_dataset)}")
 
 # 2. Training Configuration
 config = SFTConfig(
@@ -99,7 +113,6 @@ trainer = SFTTrainer(
     eval_dataset=eval_dataset,  # CRITICAL: Must provide eval_dataset when eval_strategy is enabled
     args=config,
     peft_config=peft_config,
-    dataset_text_field="messages",
 )
 
 # 5. Start Training