vocab : correctly identify LF token for GPT-2 style BPE tokenizer (gg…

…erganov#11496)
OpenBMB · Jan 30, 2025 · ffd0821 · ffd0821
1 parent 4314e56
commit ffd0821
Showing 1 changed file with 1 addition and 1 deletion.
diff --git 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()) {