diff --git a/examples/aquila/conf/train/train_aquila_7b.yaml b/examples/aquila/conf/train/train_aquila_7b.yaml
index 0c7fe2c5..1210deef 100644
--- a/examples/aquila/conf/train/train_aquila_7b.yaml
+++ b/examples/aquila/conf/train/train_aquila_7b.yaml
@@ -1,11 +1,11 @@
 system:
   tensor_model_parallel_size: 1
-  pipeline_model_parallel_size: 2
+  pipeline_model_parallel_size: 1
   disable_bias_linear: True
   use_flash_attn: True
   use_distributed_optimizer: True
   precision:
-    bf16: True
+    fp16: True
     initial_loss_scale: 522893
     min_loss_scale: 1.0
     attention_softmax_in_fp32: True
@@ -19,11 +19,11 @@ system:
     save_interval: 2000 
 
 model:
-  num_layers: 12 
-  hidden_size: 2048 
+  num_layers: 32 
+  hidden_size: 4096 
   num_attention_heads: 32 
-  seq_length: 4096 
-  max_position_embeddings: 4096
+  seq_length: 2048 
+  max_position_embeddings: 2048
   norm_epsilon: 1e-5
   use_rotary_position_embeddings: true
   no_position_embedding: true
@@ -37,10 +37,10 @@ model:
   hidden_dropout: 0.0
   weight_decay: 0.1
   clip_grad: 1.0
-  train_samples: 1000
+  train_samples: 1002539063
   eval_iters: 0
   micro_batch_size: 2
-  global_batch_size: 16
+  global_batch_size: 1728
   seed: 1234
 
   optimizer:
@@ -50,14 +50,15 @@ model:
     lr_scheduler:
       lr: 2.0e-5
       min_lr: 2.0e-6
-      lr_warmup_samples: 10
+      lr_warmup_samples: 3076172
       lr_decay_style: cosine 
 
 data:
-  data_path: ./build/data/pile_wikipedia_demo
+  data_path: ${data_path:??}
   split: 1
   tokenizer:
-    tokenizer_type: QwenTokenizerFS
-    tokenizer_path: examples/aquila/qwentokenizer
-    vocab_size: 151851
-    make_vocab_size_divisible_by: 64
\ No newline at end of file
+    tokenizer_type: AquilaTokenizerFS
+    vocab_file: ./examples/aquila/tokenizer/vocab.json
+    merge_file: ./examples/aquila/tokenizer/merges.txt
+    special_tokens_file: ./examples/aquila/tokenizer/special_tokens.txt
+    vocab_size: 100008
\ No newline at end of file