summaryrefslogtreecommitdiff
path: root/llama.cpp
diff options
context:
space:
mode:
authorJared Van Bortel <jared@nomic.ai>2023-11-15 11:34:47 -0500
committerGitHub <noreply@github.com>2023-11-15 11:34:47 -0500
commita6fc554e268634494f33b0de76f9dde650dd292f (patch)
treea4e9fe3a98858a726e046d415bb588cd245bc286 /llama.cpp
parent1cf2850d52bb027aa215e039ed9a0c61beeef8d3 (diff)
llama : restore prefix space in llama tokenizer (#4081)
Diffstat (limited to 'llama.cpp')
-rw-r--r--llama.cpp5
1 files changed, 4 insertions, 1 deletions
diff --git a/llama.cpp b/llama.cpp
index 01522fdb..92c4536c 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -6283,7 +6283,10 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
// by modifying llm_tokenizer_x to operate with string offsets like pre-tokenizer
// and passing 'add space prefix' as bool argument
//
- auto raw_text = (special ? "" : " ") + fragment.raw_text.substr(fragment.offset, fragment.length);
+ auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
+ if (&fragment == &fragment_buffer.front()) {
+ raw_text = " " + raw_text; // prefix with space if the first token is not special
+ }
#ifdef PRETOKENIZERDEBUG
fprintf(stderr,"TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());