diff --git a/downloads-generation/models_class1_pan/GENERATE.WITH_HPC_CLUSTER.sh b/downloads-generation/models_class1_pan/GENERATE.WITH_HPC_CLUSTER.sh
index 84e040cdd223494cbc39165b121ba06dc9d8a303..5d21a2e60474cccae866c821c5839133fe804f40 100755
--- a/downloads-generation/models_class1_pan/GENERATE.WITH_HPC_CLUSTER.sh
+++ b/downloads-generation/models_class1_pan/GENERATE.WITH_HPC_CLUSTER.sh
@@ -62,6 +62,8 @@ do
         --max-models 32 \
         --num-jobs $NUM_JOBS --max-tasks-per-worker 1 --gpus $GPUS --max-workers-per-gpu 1
 
+    cp "$MODELS_DIR/train_data.csv.bz2" "models.${kind}/"
+
     # Percentile rank calibration is run on the cluster.
     # For now we calibrate percentile ranks only for alleles for which there
     # is training data. Calibrating all alleles would be too slow.
diff --git a/downloads-generation/models_class1_pan/GENERATE.sh b/downloads-generation/models_class1_pan/GENERATE.sh
index 216942dcb53c2bad47b17bcd8500daab226502fa..a0842d31b226b684eeec48520a2ba7c85e32a101 100755
--- a/downloads-generation/models_class1_pan/GENERATE.sh
+++ b/downloads-generation/models_class1_pan/GENERATE.sh
@@ -55,6 +55,8 @@ do
         --num-jobs 0 \
         --num-jobs $NUM_JOBS --max-tasks-per-worker 1 --gpus $GPUS --max-workers-per-gpu 1
 
+    cp "$MODELS_DIR/train_data.csv.bz2" "models.${kind}/"
+
     # For now we calibrate percentile ranks only for alleles for which there
     # is training data. Calibrating all alleles would be too slow.
     # This could be improved though.