diff --git a/examples/train-text-from-scratch/train-text-from-scratch.cpp b/examples/train-text-from-scratch/train-text-from-scratch.cpp
index aaae5e576..76cf501a5 100644
--- a/examples/train-text-from-scratch/train-text-from-scratch.cpp
+++ b/examples/train-text-from-scratch/train-text-from-scratch.cpp
@@ -1188,19 +1188,23 @@ void save_llama_model_file(const char * filename, const char * fn_vocab_model, s
 void load_checkpoint_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model, struct ggml_opt_context * opt) {
     load_llama_model_gguf(fctx, f_ggml_ctx, model);
 
-    uint32_t file_version;
-    GGUF_GET_KEY(fctx, file_version,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_FILE_VERSION);
-    GGML_ASSERT(file_version == 0);
+    if (gguf_find_key(fctx, LLM_KV_TRAINING_FILE_VERSION) >= 0) {
+        uint32_t file_version = 0xFFFFFFFFu;
+        GGUF_GET_KEY(fctx, file_version,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_FILE_VERSION);
+        GGML_ASSERT(file_version == 0);
 
-    std::string train_type = LLM_KV_TRAINING_TYPE_TRAIN_MODEL;
-    GGUF_GET_KEY(fctx, train_type,           gguf_get_val_str, GGUF_TYPE_STRING, false, LLM_KV_TRAINING_TYPE);
-    GGML_ASSERT(train_type == LLM_KV_TRAINING_TYPE_TRAIN_MODEL);
+        std::string train_type = LLM_KV_TRAINING_TYPE_TRAIN_MODEL;
+        GGUF_GET_KEY(fctx, train_type,           gguf_get_val_str, GGUF_TYPE_STRING, false, LLM_KV_TRAINING_TYPE);
+        GGML_ASSERT(train_type == LLM_KV_TRAINING_TYPE_TRAIN_MODEL);
 
-    GGUF_GET_KEY(fctx, model->train_its,     gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_ITERATION_COUNT);
-    GGUF_GET_KEY(fctx, model->train_samples, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_SAMPLE_COUNT);
-    GGUF_GET_KEY(fctx, model->train_tokens,  gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_TOKEN_COUNT);
+        GGUF_GET_KEY(fctx, model->train_its,     gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_ITERATION_COUNT);
+        GGUF_GET_KEY(fctx, model->train_samples, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_SAMPLE_COUNT);
+        GGUF_GET_KEY(fctx, model->train_tokens,  gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_TOKEN_COUNT);
 
-    load_opt_context_gguf(fctx, f_ggml_ctx, opt);
+        load_opt_context_gguf(fctx, f_ggml_ctx, opt);
+    } else {
+        printf("%s: loaded llama model as checkpoint\n", __func__);
+    }
 }
 
 void save_checkpoint_gguf(struct gguf_context * fctx, const char * fn_vocab_model, struct my_llama_model * model, struct ggml_opt_context * opt) {