diff --git a/VinVL_MODEL_ZOO.md b/VinVL_MODEL_ZOO.md
index 7dff5fd..3b6ded5 100644
--- a/VinVL_MODEL_ZOO.md
+++ b/VinVL_MODEL_ZOO.md
@@ -236,7 +236,7 @@ Finetuned model checkpoint (w/ CIDEr optimization): [coco_captioning_base_scst.z
 1) First train with cross-entropy loss (8 V100 with 16G mem):
 ```bash
 python oscar/run_captioning.py \
-    --model_name_or_path pretrained_models/image_captioning/pretrained_base \
+    --model_name_or_path pretrained_models/image_captioning/pretrained_base/checkpoint-2000000 \
     --do_train \
     --do_lower_case \
     --add_od_labels \
@@ -275,7 +275,7 @@ Finetuned model checkpoint (w/ CIDEr optimization): [coco_captioning_large_scst.
 1) First train with cross-entropy loss (8 V100 with 32G mem):
 ```bash
 python oscar/run_captioning.py \
-    --model_name_or_path pretrained_models/image_captioning/pretrained_large \
+    --model_name_or_path pretrained_models/image_captioning/pretrained_large/checkpoint-1410000 \
     --do_train \
     --do_lower_case \
     --add_od_labels \
@@ -417,4 +417,4 @@ To set textb_sample_mode=2 for coco_flickr30k_gqa has the potential to emphasize
     --textb_sample_mode 2 --texta_false_prob 0.25 \
     --extra_dataset_file googlecc_sbu_oi_x152c4big2exp168.yaml \
     --extra_textb_sample_mode 1 --extra_loss_weight 0.5
-```
\ No newline at end of file
+```