diff --git a/README.md b/README.md
index 3271b4b9..48a02944 100644
--- a/README.md
+++ b/README.md
@@ -64,7 +64,7 @@ We include a collection of representative instruction datasets in our exploratio
 
 Please check these datasets for licenses and restrictions around their use!
 
-You can also find the processed [Tulu v1](https://huggingface.co/datasets/allenai/tulu-v1-sft-mixture) and [Tulu v2](https://huggingface.co/datasets/allenai/tulu-v2-sft-mixture) SFT datasets on HuggingFace.
+You can also find the processed [Tulu v1](https://huggingface.co/datasets/allenai/tulu-v1-sft-mixture) and [Tulu v2](https://huggingface.co/datasets/allenai/tulu-v2-sft-mixture) SFT datasets on HuggingFace. Note that the train data preparation script will not precisely recreate the Tulu v2 mixture due to randomness in the generation and shifts in data availability - see [this PR](https://github.com/allenai/open-instruct/pull/156) for some details. If you need exactly yhe training data used, the HuggingFace mixture is exactly this - the exact same data used during model training.
 
 ### Model preparation
 
diff --git a/open_instruct/reformat_datasets.py b/open_instruct/reformat_datasets.py
index fd6e1051..1d43bdaa 100644
--- a/open_instruct/reformat_datasets.py
+++ b/open_instruct/reformat_datasets.py
@@ -83,7 +83,7 @@ def convert_super_ni_data(data_dir, output_dir, zero_shot_examples_per_task=60,
 def convert_cot_data(data_dir, output_dir, num_zero_shot_examples=50000, num_few_shot_examples=50000):
     os.makedirs(output_dir, exist_ok=True)
     examples = []
-    if num_few_shot_examples > 0:
+    if num_zero_shot_examples > 0:
         with open(os.path.join(data_dir, "cot_zsopt.jsonl"), "r") as fin:
             zero_shot_examples = [json.loads(line) for line in fin]
             if num_zero_shot_examples < len(zero_shot_examples):
@@ -493,8 +493,14 @@ def convert_lima_data(data_dir, output_dir, num_examples=None):
 def convert_wizardlm_data(data_dir, output_dir, num_examples=30000):
     os.makedirs(output_dir, exist_ok=True)
     examples = []
-    with open(os.path.join(data_dir, "WizardLM_evol_instruct_V2_143k.json"), "r") as fin:
-        examples = json.load(fin)
+    # check if the original json file exists
+    if os.path.exists(os.path.join(data_dir, "WizardLM_evol_instruct_V2_143k.json")):
+        # proceed as normally
+        with open(os.path.join(data_dir, "WizardLM_evol_instruct_V2_143k.json"), "r") as fin:
+            examples = json.load(fin)
+    else: # try other data train-00000-of-00001-004cd1ba9dc05e6c.parquet
+        df = pd.read_parquet(os.path.join(data_dir, "train-00000-of-00001-004cd1ba9dc05e6c.parquet"))
+        examples = df.to_dict(orient="records")
     if num_examples:
         examples = random.sample(examples, k=num_examples)
 
diff --git a/scripts/prepare_train_data.sh b/scripts/prepare_train_data.sh
index 18ae4f03..5e45e872 100755
--- a/scripts/prepare_train_data.sh
+++ b/scripts/prepare_train_data.sh
@@ -1,5 +1,7 @@
-# check if there is $HF_TOKEN in the environment variables
-if [ -z "$HF_TOKEN" ]
+# check if there is $HF_TOKEN in the environment variables, or if the huggingface-cli is installed and logged in 
+if huggingface-cli whoami &>/dev/null; then
+    echo "Logged in to HuggingFace."
+elif [ -z "$HF_TOKEN" ]
 then
     echo "Warning: HuggingFace dataset LIMA requires permissive access."
     echo "Warning: Please request the access at https://huggingface.co/datasets/GAIR/lima and set the HF_TOKEN environment variable before running this script."
@@ -77,13 +79,16 @@ python scripts/split_sharegpt_conversations.py \
     --model-name-or-path oobabooga/llama-tokenizer \
     --max-length 4096
 
-
 echo "Downloading LIMA dataset..."
-wget --header="Authorization: Bearer $HF_TOKEN" -P data/raw_train/lima/ https://huggingface.co/datasets/GAIR/lima/raw/main/train.jsonl
-
+if huggingface-cli whoami &>/dev/null; then
+    huggingface-cli download GAIR/lima --repo-type dataset --local-dir data/raw_train/lima/
+else
+    wget --header="Authorization: Bearer $HF_TOKEN" -P data/raw_train/lima/ https://huggingface.co/datasets/GAIR/lima/raw/main/train.jsonl
+fi
 
 echo "Downloading WizardLM dataset..."
-wget -P data/raw_train/wizardlm/ https://huggingface.co/datasets/WizardLM/WizardLM_evol_instruct_V2_196k/resolve/main/WizardLM_evol_instruct_V2_143k.json
+# original data removed wget -P data/raw_train/wizardlm/ https://huggingface.co/datasets/WizardLM/WizardLM_evol_instruct_V2_196k/resolve/main/WizardLM_evol_instruct_V2_143k.json
+wget -P data/raw_train/wizardlm/ https://huggingface.co/datasets/ai2-adapt-dev/wizardlm-backup/resolve/main/data/train-00000-of-00001.parquet?download=true
 
 
 echo "Downloading the OpenOrca dataset..."