update examples

Former-commit-id: 3b5f138155d96b346bda18e465cf60ec7d99e19c

examples/full_multi_gpu/llama3_full_predict.yaml

@@ -1,12 +1,12 @@
-# model
+### model
 model_name_or_path: saves/llama3-8b/full/sft
 
-# method
+### method
 stage: sft
 do_predict: true
 finetuning_type: full
 
-# dataset
+### dataset
 dataset: identity,alpaca_gpt4_en
 template: llama3
 cutoff_len: 1024
@@ -14,10 +14,10 @@ max_samples: 50
 overwrite_cache: true
 preprocessing_num_workers: 16
 
-# output
+### output
 output_dir: saves/llama3-8b/full/predict
 overwrite_output_dir: true
 
-# eval
+### eval
 per_device_eval_batch_size: 1
 predict_with_generate: true

examples/full_multi_gpu/llama3_full_sft.yaml

@@ -1,16 +1,16 @@
-# model
+### model
 model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
 
-# method
+### method
 stage: sft
 do_train: true
 finetuning_type: full
 
-# ddp
+### ddp
 ddp_timeout: 180000000
 deepspeed: examples/deepspeed/ds_z3_config.json
 
-# dataset
+### dataset
 dataset: identity,alpaca_gpt4_en
 template: llama3
 cutoff_len: 1024
@@ -18,14 +18,14 @@ max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
 
-# output
+### output
 output_dir: saves/llama3-8b/full/sft
 logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
 
-# train
+### train
 per_device_train_batch_size: 1
 gradient_accumulation_steps: 2
 learning_rate: 0.0001
@@ -34,7 +34,7 @@ lr_scheduler_type: cosine
 warmup_steps: 0.1
 fp16: true
 
-# eval
+### eval
 val_size: 0.1
 per_device_eval_batch_size: 1
 evaluation_strategy: steps