From 93a4bb8b106261c2734d25d3d0e0e85e451bbad0 Mon Sep 17 00:00:00 2001
From: Weizhi Wang <wangweizhihygmail.com>
Date: Sat, 5 Aug 2023 07:13:15 -0400
Subject: [PATCH] fix path bug in training script

---
 train_scripts/train_longmem.sh | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/train_scripts/train_longmem.sh b/train_scripts/train_longmem.sh
index 7b1f034..7e27a79 100644
--- a/train_scripts/train_longmem.sh
+++ b/train_scripts/train_longmem.sh
@@ -13,7 +13,7 @@ fairseq-train ${DATA_DIR}  \
     --save-interval-updates 10000 --sample-break-mode none \
     --tokens-per-sample 1024 \
     --batch-size 8 --total-num-update 100000 --seed 42 \
-    --pretrained-model-path bigscience/bloom-1b7 \
+    --pretrained-model-path /path/to/gpt2_pretrained_model \
     --layer-reduction-factor 2 \
     --disable-validation \
     --use-external-memory --memory-size 65536 \
@@ -23,4 +23,6 @@ fairseq-train ${DATA_DIR}  \
     --no-token-positional-embeddings \
     --data-no-shuffle \
     --retrieval-layer-index 17 \
-    --reload-ptm-layer
\ No newline at end of file
+    --reload-ptm-layer
+
+# The --pre-trained-model path refers to the path to reproduced GPT-2-Medium checkpoints. You can find the downloading Google Drive url in README.
\ No newline at end of file