diff --git a/dev/data/hellaswag.py b/dev/data/hellaswag.py
index 4640166..f4395e9 100644
--- a/dev/data/hellaswag.py
+++ b/dev/data/hellaswag.py
@@ -101,16 +101,11 @@ def render_example(example):
 
 def iterate_examples(split):
     # there are 10,042 examples in total in val
-    n = 0
     download(split)
     with open(os.path.join(DATA_CACHE_DIR, f"hellaswag_{split}.jsonl"), "r") as f:
         for line in f:
             example = json.loads(line)
-            n += 1
             yield example
-            # DEBUGGING, TODO REMOVE
-            if n >= 101:
-                break
 
 @torch.no_grad()
 def evaluate(model_type, device):
diff --git a/train_gpt2.cu b/train_gpt2.cu
index 11de22f..62e701a 100644
--- a/train_gpt2.cu
+++ b/train_gpt2.cu
@@ -2694,7 +2694,12 @@ int main(int argc, char *argv[]) {
     // build an EvalLoader for HellaSwag
     EvalLoader eval_loader;
     const char* hellaswag_path = "dev/data/hellaswag/hellaswag_val.bin";
-    evalloader_init(&eval_loader, hellaswag_path, B, T, multi_gpu_config.process_rank, multi_gpu_config.num_processes);
+    const char hellaswag_available = access(hellaswag_path, F_OK) == 0;
+    if (hellaswag_available) {
+        evalloader_init(&eval_loader, hellaswag_path, B, T, multi_gpu_config.process_rank, multi_gpu_config.num_processes);
+    }
+    printf0("| hellaswag available   | %-50s |\n", hellaswag_available ? "yes" : "no");
+    printf0("+-----------------------+----------------------------------------------------+\n");
 
     // pretty print in a table the multi-gpu configuration as well
     set_zero_configs(&multi_gpu_config, zero_stage, model.num_parameters);
@@ -2702,6 +2707,11 @@ int main(int argc, char *argv[]) {
     printf0("| zero_stage            | %-50d |\n", multi_gpu_config.zero_stage);
     printf0("+-----------------------+----------------------------------------------------+\n");
 
+    // prints outside of pretty table to here and below
+    if (!hellaswag_available) {
+        printf0("HellaSwag eval not found at %s, skipping its evaluation\n", hellaswag_path);
+        printf0("You can run `python dev/data/hellaswag.py` to export and use it.\n");
+    }
     // more prints related to allocations from gpt2_build_from_checkpoint down here to not mess up our table above
     printf0("num_parameters: %zu => bytes: %zu\n", model.num_parameters, model.num_parameters_bytes);
     printf0("allocated %d MiB for model parameters\n", (int)round(model.num_parameters_bytes / (1024 * 1024)));
@@ -2755,11 +2765,13 @@ int main(int argc, char *argv[]) {
         }
 
         // once in a while estimate HellaSwag accuracy
-        if (step % val_loss_every == 0 || last_step) {
+        if (hellaswag_available &&
+           (step % val_loss_every == 0 || last_step)) {
             NvtxRange evaluation_range("evaluation");
             float eval_acc_norm = 0.0f;
             evalloader_reset(&eval_loader);
             for (int i = 0; i < eval_loader.num_batches; i++) {
+                if (i % 10 == 0) { printf("evaluating HellaSwag: %d/%d\r", i, eval_loader.num_batches); }
                 evalloader_next_batch(&eval_loader);
                 gpt2_forward(&model, eval_loader.inputs, eval_loader.targets, B, T);
                 int correct = evalloader_stat_losses(&eval_loader, model.cpu_losses_fp32);