MCPcopy Create free account
hub / github.com/0xShug0/audio.cpp / unicode_regex_split_custom_llama3

Function unicode_regex_split_custom_llama3

external/llama_tokenizer/unicode.cpp:333–471  ·  view source on GitHub ↗

LLAMA3 system regex: "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"

Source from the content-addressed store, hash-verified

331
332// LLAMA3 system regex: "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"
333static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string & text, const std::vector<size_t> & offsets) {
334 std::vector<size_t> bpe_offsets; // store the offset of each word
335 bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
336
337 const auto cpts = unicode_cpts_from_utf8(text);
338
339 size_t start = 0;
340 for (auto offset : offsets) {
341 const size_t offset_ini = start;
342 const size_t offset_end = start + offset;
343 assert(offset_end <= cpts.size());
344 start = offset_end;
345
346 static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;
347 auto _get_cpt = [&] (const size_t pos) -> uint32_t {
348 return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
349 };
350
351 auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
352 return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
353 };
354
355 size_t _prev_end = offset_ini;
356 auto _add_token = [&] (const size_t end) -> size_t {
357 assert(_prev_end <= end && end <= offset_end);
358 size_t len = end - _prev_end;
359 if (len > 0) {
360 bpe_offsets.push_back(len);
361 }
362 _prev_end = end;
363 //if (len > 0) {
364 // std::string s = "";
365 // for(size_t p = end-len; p < end; p++)
366 // s += unicode_cpt_to_utf8(cpts[p]);
367 // printf(">>> '%s'\n", s.c_str());
368 //}
369 return len;
370 };
371
372 for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
373 const uint32_t cpt = _get_cpt(pos);
374 const auto flags = _get_flags(pos);
375
376 // regex: (?i:'s|'t|'re|'ve|'m|'ll|'d) // case insensitive
377 if (cpt == '\'' && pos+1 < offset_end) {
378 uint32_t cpt_next = unicode_tolower(_get_cpt(pos+1));
379 if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
380 pos += _add_token(pos+2);
381 continue;
382 }
383 if (pos+2 < offset_end) {
384 uint32_t cpt_next_next = unicode_tolower(_get_cpt(pos+2));
385 if ((cpt_next == 'r' && cpt_next_next == 'e') ||
386 (cpt_next == 'v' && cpt_next_next == 'e') ||
387 (cpt_next == 'l' && cpt_next_next == 'l')) {
388 pos += _add_token(pos+3);
389 continue;
390 }

Callers 1

Calls 5

unicode_cpts_from_utf8Function · 0.85
unicode_tolowerFunction · 0.85
as_uintMethod · 0.80
sizeMethod · 0.45

Tested by

no test coverage detected