Upload train_survival.py with huggingface_hub

train_survival.py

@@ -17,7 +17,7 @@ OUTPUT_MODEL_ID = "sunkencity/survival-expert-3b"
 # Load Dataset
 dataset = load_dataset(DATASET_ID, split="train")
 
-# Load Model with Quantization (for efficiency)
+# Load Model
 bnb_config = BitsAndBytesConfig(
     load_in_4bit=True,
     bnb_4bit_quant_type="nf4",
@@ -33,7 +33,7 @@ model = AutoModelForCausalLM.from_pretrained(
 tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
 tokenizer.pad_token = tokenizer.eos_token
 
-# LoRA Configuration
+# LoRA
 peft_config = LoraConfig(
     r=16,
     lora_alpha=32,
@@ -43,8 +43,7 @@ peft_config = LoraConfig(
     target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
 )
 
-# Training Arguments
-# Removed max_seq_length from SFTConfig as it caused a TypeError
+# Args
 training_args = SFTConfig(
     output_dir="./results",
     num_train_epochs=3,
@@ -55,32 +54,28 @@ training_args = SFTConfig(
     push_to_hub=True,
     hub_model_id=OUTPUT_MODEL_ID,
     fp16=True,
-    dataset_text_field="text", 
+    dataset_text_field="text",
     packing=False
 )
 
-# Formatting function for SFT (Chat format)
 def formatting_prompts_func(example):
     output_texts = []
     for i in range(len(example['instruction'])):
         instruction = example['instruction'][i]
         response = example['response'][i]
-        
-        # Qwen/Llama chat template format
         text = f"<|im_start|>user\n{instruction}<|im_end|>\n<|im_start|>assistant\n{response}<|im_end|>"
         output_texts.append(text)
     return output_texts
 
 # Trainer
-# max_seq_length is passed to SFTTrainer directly
 trainer = SFTTrainer(
     model=model,
     train_dataset=dataset,
     peft_config=peft_config,
     formatting_func=formatting_prompts_func,
     args=training_args,
-    tokenizer=tokenizer,
-    max_seq_length=1024 
+    processing_class=tokenizer, # New name for tokenizer
+    max_seq_length=1024 # Passed here
 )
 
 print("Starting training...")