Upload 2 files

dataset_info.json

@@ -0,0 +1,22 @@
+{
+	"train_sample": {
+		"file_name": "train_sample.json",
+		"columns": {
+			"prompt": "instruction",
+			"query": "input",
+			"response": "output",
+			"system": "system",
+			"history": "history"
+		}
+	},
+	"test_sample": {
+		"file_name": "train_sample.json",
+		"columns": {
+			"prompt": "instruction",
+			"query": "input",
+			"response": "output",
+			"system": "system",
+			"history": "history"
+		}
+	}
+}

llama3_guardrail_lora_sft.yaml

@@ -0,0 +1,37 @@
+### examples/train_lora/llama3_lora_sft.yaml
+model_name_or_path: "meta-llama/Llama-Guard-3-1B"
+
+stage: sft
+do_train: true
+do_eval: true
+finetuning_type: lora
+lora_target: all
+
+dataset: train_sample
+eval_dataset: test_sample
+dataset_dir: ./data
+template: llama3
+cutoff_len: 4096
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 64
+
+output_dir: ./saves/llama3-1b/lora/sft
+logging_steps: 1
+save_steps: 10
+plot_loss: true
+overwrite_output_dir: true
+
+per_device_train_batch_size: 4
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+# val_size: 0.1
+per_device_eval_batch_size: 16
+eval_strategy: steps
+eval_steps: 1