vocab : correctly identify LF token for GPT-2 style BPE tokenizer (#11496)

diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
index 561f8bdb840106a871d6866effe8d71c2030dd26..ad9ffe66aa749d97ec19ebe7286ed226e63b5575 100644 (file)
--- a/src/llama-vocab.cpp
+++ b/src/llama-vocab.cpp
@@ -1692,7 +1692,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
         GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
         linefeed_id = ids[0];
     } else {
-        const std::vector<int> ids = tokenize("\xC4\x8A", false); // U+010A
+        const std::vector<int> ids = tokenize("\n", false);
 
         //GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
         if (ids.empty()) {