Skip to content

armgit / laya

Public Imported from GitHub
0

Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.

fix(java): the pre-tokenizer classified 9,917 code points by the JDK's Unicode, not the reference's

Adeshina Lasisi0b4ebcab936a
diff --git a/.github/workflows/java.yml b/.github/workflows/java.yml
index 9376a859..370671d7 100644
--- a/.github/workflows/java.yml
+++ b/.github/workflows/java.yml
@@ -99,6 +99,8 @@ jobs:
- name: Language tables match laya/lang.py
if: ${{ !cancelled() }}
run: python laya-java/scripts/gen_language_tables.py --check
+ # The pre-tokenizer's letter/number classes are probed from the reference tokenizer, so this
+ # one needs a checkpoint and runs in the parity lane instead -- see that job.
- uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
if: ${{ failure() }}
@@ -165,6 +167,10 @@ jobs:
LAYA_ONNX_GRAPH: ${{ github.workspace }}/laya-java/.work/onnx/multilingual/laya.onnx
LAYA_PREDICT_MODEL: multilingual
run: python laya-java/scripts/gen_fixtures.py --strict
+ - name: Pre-tokenizer class tables match the reference
+ run: >
+ python laya-java/scripts/gen_pretokenizer_tables.py --check
+ --tokenizer laya-java/.work/checkpoints/english/tokenizer/tokenizer.json
# Informational: float noise across platforms is expected in the model-backed family, and the
# Java test's tolerance is the gate rather than byte equality of the file.
- name: Fixture drift vs. committed (informational)
diff --git a/laya-java/laya-java/src/main/java/com/convaiinnovations/laya/tokenizer/PreTokenizer.java b/laya-java/laya-java/src/main/java/com/convaiinnovations/laya/tokenizer/PreTokenizer.java
index eb266b08..b77de0a2 100644
--- a/laya-java/laya-java/src/main/java/com/convaiinnovations/laya/tokenizer/PreTokenizer.java
+++ b/laya-java/laya-java/src/main/java/com/convaiinnovations/laya/tokenizer/PreTokenizer.java
@@ -93,25 +93,28 @@ abstract class PreTokenizer {
/** GPT-2 byte-level pre-tokenization: regex cut, then UTF-8 bytes into a printable alphabet. */
static final class ByteLevel extends PreTokenizer {
- /**
- * The GPT-2 cut. Ordered alternation, and the order is the specification: {@code \s+(?!\S)}
- * must be tried before {@code \s+} so that a run of whitespace followed by a non-space
- * leaves its last space to the following word, which is what produces {@code "Ġthere"}
- * rather than a lone whitespace token.
+ /*
+ * The GPT-2 cut, as a scanner rather than a regex.
+ *
+ * The pattern is
+ * 's|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+
+ * and the ORDER is the specification. `\s+(?!\S)` must be tried before `\s+` so that a
+ * run of whitespace followed by a non-space leaves its last space to the following word,
+ * which is what produces "Ġthere" rather than a lone whitespace token.
*
- * <p>{@link Pattern#UNICODE_CHARACTER_CLASS} is not optional: Java's {@code \s} is
- * ASCII-only by default, where the Rust implementation's is the Unicode
- * {@code White_Space} property. Without the flag, a non-breaking space or an ideographic
- * space would fall through to the punctuation branch and tokenize differently.
+ * WHY NOT java.util.regex. `\p{L}` and `\p{N}` follow the JDK's Unicode version, and the
+ * reference follows its own -- Corretto 17 ships Unicode 13.0 while the reference
+ * classifies Unicode 15.1 and 16.0 additions as letters. They disagree on 9,917 code
+ * points, and for every one of them the seven contraction tokens could never form: when
+ * ` ?\p{L}+` fails, ` ?[^\s\p{L}\p{N}]+` swallows the apostrophe into the same piece.
+ * "The X's value" tokenized differently for 9,917 values of X on the shipped english
+ * checkpoint, and the same jar gave different ids on a different JDK -- the opposite of
+ * what this port is for. The classes now come from PreTokenizerTables, probed from the
+ * reference. Whitespace is Unicode.isWhiteSpace, which matches Rust's
+ * char::is_whitespace on every code point and is what UNICODE_CHARACTER_CLASS bought.
*/
- private static final Pattern GPT2 = Pattern.compile(
- "'s|'t|'re|'ve|'m|'ll|'d"
- + "| ?\\p{L}+"
- + "| ?\\p{N}+"
- + "| ?[^\\s\\p{L}\\p{N}]+"
- + "|\\s+(?!\\S)"
- + "|\\s+",
- Pattern.UNICODE_CHARACTER_CLASS);
+ /** The seven contraction literals, in the pattern's own order. */
+ private static final String[] CONTRACTIONS = {"'s", "'t", "'re", "'ve", "'m", "'ll", "'d"};
/** byte value -> the character that stands for it. */
private static final char[] BYTE_TO_CHAR = buildByteToChar();
@@ -200,15 +203,103 @@ abstract class PreTokenizer {
}
return out;
}
- Matcher matcher = GPT2.matcher(source);
- while (matcher.find()) {
- if (matcher.end() > matcher.start()) {
- out.add(encode(source.substring(matcher.start(), matcher.end())));
+ int at = 0;
+ while (at < source.length()) {
+ int end = cut(source, at);
+ if (end > at) {
+ out.add(encode(source.substring(at, end)));
+ at = end;
+ } else {
+ // Unreachable: every branch of the pattern consumes at least one character
+ // for any input, and the final `\s+` catches whitespace. Guarded anyway,
+ // because a scanner that returns its own position loops forever.
+ throw new IllegalStateException(
+ "the pre-tokenizer made no progress at offset " + at);
}
}
return out;
}
+ /**
+ * The end of the piece beginning at {@code from}, following the pattern's alternation in
+ * its own order.
+ *
+ * @return an offset strictly greater than {@code from}
+ */
+ private static int cut(String source, int from) {
+ int length = source.length();
+ // 's|'t|'re|'ve|'m|'ll|'d -- literals, tried first and in this order.
+ for (String contraction : CONTRACTIONS) {
+ if (source.startsWith(contraction, from)) {
+ return from + contraction.length();
+ }
+ }
+ // ` ?\p{L}+`, ` ?\p{N}+`, ` ?[^\s\p{L}\p{N}]+`: an optional single leading space,
+ // then a run of one class. The optional space is only taken when a character of that
+ // class follows it, which is what the regex's backtracking amounts to here.
+ for (int branch = 0; branch < 3; branch++) {
+ int start = from;
+ if (source.charAt(start) == ' ') {
+ int after = start + 1;
+ if (after < length && inClass(source, after, branch)) {
+ return runEnd(source, after, branch);
+ }
+ } else if (inClass(source, start, branch)) {
+ return runEnd(source, start, branch);
+ }
+ }
+ // `\s+(?!\S)` then `\s+`. Greedy, then backtracked so that the match is not
+ // followed by a non-space: for a run followed by a word, that leaves the run's LAST
+ // space to the following word, which is what makes "Ġthere" rather than a lone
+ // whitespace piece. A run that reaches the end of the input is taken whole.
+ if (Unicode.isWhiteSpace(source.codePointAt(from))) {
+ int end = from;
+ while (end < length) {
+ int cp = source.codePointAt(end);
+ if (!Unicode.isWhiteSpace(cp)) {
+ break;
+ }
+ end += Character.charCount(cp);
+ }
+ if (end >= length) {
+ return end;
+ }
+ int backtracked = end - Character.charCount(source.codePointBefore(end));
+ if (backtracked > from) {
+ return backtracked;
+ }
+ return end; // the single-space case: `\s+(?!\S)` fails, `\s+` wins
+ }
+ // Unreachable for any input: the three class branches partition every non-whitespace
+ // code point. A single code point is consumed rather than returning `from`, because a
+ // scanner that makes no progress loops forever.
+ return from + Character.charCount(source.codePointAt(from));
+ }
+
+ /** Whether the code point at {@code at} belongs to branch 0 letter, 1 number, 2 other. */
+ private static boolean inClass(String source, int at, int branch) {
+ int cp = source.codePointAt(at);
+ switch (branch) {
+ case 0:
+ return PreTokenizerTables.isLetter(cp);
+ case 1:
+ return PreTokenizerTables.isNumber(cp);
+ default:
+ return !Unicode.isWhiteSpace(cp)
+ && !PreTokenizerTables.isLetter(cp)
+ && !PreTokenizerTables.isNumber(cp);
+ }
+ }
+
+ /** The end of the maximal run of {@code branch}'s class starting at {@code at}. */
+ private static int runEnd(String source, int at, int branch) {
+ int end = at;
+ while (end < source.length() && inClass(source, end, branch)) {
+ end += Character.charCount(source.codePointAt(end));
+ }
+ return end;
+ }
+
/** One piece's UTF-8 bytes, each replaced by its stand-in character. */
private static String encode(String piece) {
byte[] bytes = piece.getBytes(StandardCharsets.UTF_8);
diff --git a/laya-java/laya-java/src/main/java/com/convaiinnovations/laya/tokenizer/PreTokenizerTables.java b/laya-java/laya-java/src/main/java/com/convaiinnovations/laya/tokenizer/PreTokenizerTables.java
new file mode 100644
index 00000000..97f3ec83
--- /dev/null
+++ b/laya-java/laya-java/src/main/java/com/convaiinnovations/laya/tokenizer/PreTokenizerTables.java
@@ -0,0 +1,232 @@
+// GENERATED by laya-java/scripts/gen_pretokenizer_tables.py -- do not edit.
+//
+// Regenerate with:
+// python laya-java/scripts/gen_pretokenizer_tables.py --tokenizer <path/to/tokenizer.json>
+// and `--check` fails when the committed copy no longer matches the reference.
+package com.convaiinnovations.laya.tokenizer;
+
+/**
+ * The letter and number classes the GPT-2 pre-tokenizer pattern splits on, recorded from the
+ * reference tokenizer (tokenizers 0.23.2).
+ *
+ * <p>Compiled in rather than taken from {@code java.util.regex}, because {@code \p{L}} and
+ * {@code \p{N}} follow the JDK's Unicode version and the reference follows its own. Corretto 17
+ * ships Unicode 13.0; the reference classifies Unicode 15.1 and 16.0 additions as letters. The
+ * two disagree on 9917 code points, and for every one of them the seven contraction tokens
+ * ({@code 's 't 're 've 'm 'll 'd}) could never form: when {@code ?\p{L}+} fails, the
+ * {@code ?[^\s\p{L}\p{N}]+} branch swallows the apostrophe into the same piece. "The X's value"
+ * therefore tokenized differently for 9917 values of X on the shipped english checkpoint -- and the
+ * same jar gave different ids on a different JDK, which is the opposite of what a parity port is
+ * for.
+ */
+public final class PreTokenizerTables {
+
+ /** The reference these classes were probed from. */
+ public static final String REFERENCE = "tokenizers 0.23.2";
+
+ private PreTokenizerTables() {
+ }
+
+ /**
+ * The pattern's {@code \\p{L}}: what the reference treats as a letter.
+ *
+ * <p>777 ranges covering 142963 code points, probed from
+ * the reference tokenizer.
+ */
+ private static final int[] LETTER = letterTable();
+
+ private static int[] letterTable() {
+ return new int[] {
+ 9, 13, 48, 57, 65, 90, 97, 122, 133, 133, 160, 160, 170, 170, 178, 179, 181, 181,
+ 185, 186, 188, 190, 192, 214, 216, 246, 248, 705, 710, 721, 736, 740, 748, 748,
+ 750, 750, 880, 884, 886, 887, 890, 893, 895, 895, 902, 902, 904, 906, 908, 908,
+ 910, 929, 931, 1013, 1015, 1153, 1162, 1327, 1329, 1366, 1369, 1369, 1376, 1416,
+ 1488, 1514, 1519, 1522, 1568, 1610, 1632, 1641, 1646, 1647, 1649, 1747, 1749, 1749,
+ 1765, 1766, 1774, 1788, 1791, 1791, 1808, 1808, 1810, 1839, 1869, 1957, 1969, 1969,
+ 1984, 2026, 2036, 2037, 2042, 2042, 2048, 2069, 2074, 2074, 2084, 2084, 2088, 2088,
+ 2112, 2136, 2144, 2154, 2160, 2183, 2185, 2190, 2208, 2249, 2308, 2361, 2365, 2365,
+ 2384, 2384, 2392, 2401, 2406, 2415, 2417, 2432, 2437, 2444, 2447, 2448, 2451, 2472,
+ 2474, 2480, 2482, 2482, 2486, 2489, 2493, 2493, 2510, 2510, 2524, 2525, 2527, 2529,
+ 2534, 2545, 2548, 2553, 2556, 2556, 2565, 2570, 2575, 2576, 2579, 2600, 2602, 2608,
+ 2610, 2611, 2613, 2614, 2616, 2617, 2649, 2652, 2654, 2654, 2662, 2671, 2674, 2676,
+ 2693, 2701, 2703, 2705, 2707, 2728, 2730, 2736, 2738, 2739, 2741, 2745, 2749, 2749,
+ 2768, 2768, 2784, 2785, 2790, 2799, 2809, 2809, 2821, 2828, 2831, 2832, 2835, 2856,
+ 2858, 2864, 2866, 2867, 2869, 2873, 2877, 2877, 2908, 2909, 2911, 2913, 2918, 2927,
+ 2929, 2935, 2947, 2947, 2949, 2954, 2958, 2960, 2962, 2965, 2969, 2970, 2972, 2972,
+ 2974, 2975, 2979, 2980, 2984, 2986, 2990, 3001, 3024, 3024, 3046, 3058, 3077, 3084,
+ 3086, 3088, 3090, 3112, 3114, 3129, 3133, 3133, 3160, 3162, 3165, 3165, 3168, 3169,
+ 3174, 3183, 3192, 3198, 3200, 3200, 3205, 3212, 3214, 3216, 3218, 3240, 3242, 3251,
+ 3253, 3257, 3261, 3261, 3293, 3294, 3296, 3297, 3302, 3311, 3313, 3314, 3332, 3340,
+ 3342, 3344, 3346, 3386, 3389, 3389, 3406, 3406, 3412, 3414, 3416, 3425, 3430, 3448,
+ 3450, 3455, 3461, 3478, 3482, 3505, 3507, 3515, 3517, 3517, 3520, 3526, 3558, 3567,
+ 3585, 3632, 3634, 3635, 3648, 3654, 3664, 3673, 3713, 3714, 3716, 3716, 3718, 3722,
+ 3724, 3747, 3749, 3749, 3751, 3760, 3762, 3763, 3773, 3773, 3776, 3780, 3782, 3782,
+ 3792, 3801, 3804, 3807, 3840, 3840, 3872, 3891, 3904, 3911, 3913, 3948, 3976, 3980,
+ 4096, 4138, 4159, 4169, 4176, 4181, 4186, 4189, 4193, 4193, 4197, 4198, 4206, 4208,
+ 4213, 4225, 4238, 4238, 4240, 4249, 4256, 4293, 4295, 4295, 4301, 4301, 4304, 4346,
+ 4348, 4680, 4682, 4685, 4688, 4694, 4696, 4696, 4698, 4701, 4704, 4744, 4746, 4749,
+ 4752, 4784, 4786, 4789, 4792, 4798, 4800, 4800, 4802, 4805, 4808, 4822, 4824, 4880,
+ 4882, 4885, 4888, 4954, 4969, 4988, 4992, 5007, 5024, 5109, 5112, 5117, 5121, 5740,
+ 5743, 5786, 5792, 5866, 5870, 5880, 5888, 5905, 5919, 5937, 5952, 5969, 5984, 5996,
+ 5998, 6000, 6016, 6067, 6103, 6103, 6108, 6108, 6112, 6121, 6128, 6137, 6160, 6169,
+ 6176, 6264, 6272, 6276, 6279, 6312, 6314, 6314, 6320, 6389, 6400, 6430, 6470, 6509,
+ 6512, 6516, 6528, 6571, 6576, 6601, 6608, 6618, 6656, 6678, 6688, 6740, 6784, 6793,
+ 6800, 6809, 6823, 6823, 6917, 6963, 6981, 6988, 6992, 7001, 7043, 7072, 7086, 7141,
+ 7168, 7203, 7232, 7241, 7245, 7293, 7296, 7306, 7312, 7354, 7357, 7359, 7401, 7404,
+ 7406, 7411, 7413, 7414, 7418, 7418, 7424, 7615, 7680, 7957, 7960, 7965, 7968, 8005,
+ 8008, 8013, 8016, 8023, 8025, 8025, 8027, 8027, 8029, 8029, 8031, 8061, 8064, 8116,
+ 8118, 8124, 8126, 8126, 8130, 8132, 8134, 8140, 8144, 8147, 8150, 8155, 8160, 8172,
+ 8178, 8180, 8182, 8188, 8192, 8202, 8232, 8233, 8239, 8239, 8287, 8287, 8304, 8305,
+ 8308, 8313, 8319, 8329, 8336, 8348, 8450, 8450, 8455, 8455, 8458, 8467, 8469, 8469,
+ 8473, 8477, 8484, 8484, 8486, 8486, 8488, 8488, 8490, 8493, 8495, 8505, 8508, 8511,
+ 8517, 8521, 8526, 8526, 8528, 8585, 9312, 9371, 9450, 9471, 10102, 10131, 11264,
+ 11492, 11499, 11502, 11506, 11507, 11517, 11517, 11520, 11557, 11559, 11559, 11565,
+ 11565, 11568, 11623, 11631, 11631, 11648, 11670, 11680, 11686, 11688, 11694, 11696,
+ 11702, 11704, 11710, 11712, 11718, 11720, 11726, 11728, 11734, 11736, 11742, 11823,
+ 11823, 12288, 12288, 12293, 12295, 12321, 12329, 12337, 12341, 12344, 12348, 12353,
+ 12438, 12445, 12447, 12449, 12538, 12540, 12543, 12549, 12591, 12593, 12686, 12690,
+ 12693, 12704, 12735, 12784, 12799, 12832, 12841, 12872, 12879, 12881, 12895, 12928,
+ 12937, 12977, 12991, 13312, 19903, 19968, 42124, 42192, 42237, 42240, 42508, 42512,
+ 42539, 42560, 42606, 42623, 42653, 42656, 42735, 42775, 42783, 42786, 42888, 42891,
+ 42957, 42960, 42961, 42963, 42963, 42965, 42972, 42994, 43009, 43011, 43013, 43015,
+ 43018, 43020, 43042, 43056, 43061, 43072, 43123, 43138, 43187, 43216, 43225, 43250,
+ 43255, 43259, 43259, 43261, 43262, 43264, 43301, 43312, 43334, 43360, 43388, 43396,
+ 43442, 43471, 43481, 43488, 43492, 43494, 43518, 43520, 43560, 43584, 43586, 43588,
+ 43595, 43600, 43609, 43616, 43638, 43642, 43642, 43646, 43695, 43697, 43697, 43701,
+ 43702, 43705, 43709, 43712, 43712, 43714, 43714, 43739, 43741, 43744, 43754, 43762,
+ 43764, 43777, 43782, 43785, 43790, 43793, 43798, 43808, 43814, 43816, 43822, 43824,
+ 43866, 43868, 43881, 43888, 44002, 44016, 44025, 44032, 55203, 55216, 55238, 55243,
+ 55291, 63744, 64109, 64112, 64217, 64256, 64262, 64275, 64279, 64285, 64285, 64287,
+ 64296, 64298, 64310, 64312, 64316, 64318, 64318, 64320, 64321, 64323, 64324, 64326,
+ 64433, 64467, 64829, 64848, 64911, 64914, 64967, 65008, 65019, 65136, 65140, 65142,
+ 65276, 65296, 65305, 65313, 65338, 65345, 65370, 65382, 65470, 65474, 65479, 65482,
+ 65487, 65490, 65495, 65498, 65500, 65536, 65547, 65549, 65574, 65576, 65594, 65596,
+ 65597, 65599, 65613, 65616, 65629, 65664, 65786, 65799, 65843, 65856, 65912, 65930,
+ 65931, 66176, 66204, 66208, 66256, 66273, 66299, 66304, 66339, 66349, 66378, 66384,
+ 66421, 66432, 66461, 66464, 66499, 66504, 66511, 66513, 66517, 66560, 66717, 66720,
+ 66729, 66736, 66771, 66776, 66811, 66816, 66855, 66864, 66915, 66928, 66938, 66940,
+ 66954, 66956, 66962, 66964, 66965, 66967, 66977, 66979, 66993, 66995, 67001, 67003,
+ 67004, 67008, 67059, 67072, 67382, 67392, 67413, 67424, 67431, 67456, 67461, 67463,
+ 67504, 67506, 67514, 67584, 67589, 67592, 67592, 67594, 67637, 67639, 67640, 67644,
+ 67644, 67647, 67669, 67672, 67702, 67705, 67742, 67751, 67759, 67808, 67826, 67828,
+ 67829, 67835, 67867, 67872, 67897, 67968, 68023, 68028, 68047, 68050, 68096, 68112,
+ 68115, 68117, 68119, 68121, 68149, 68160, 68168, 68192, 68222, 68224, 68255, 68288,
+ 68295, 68297, 68324, 68331, 68335, 68352, 68405, 68416, 68437, 68440, 68466, 68472,
+ 68497, 68521, 68527, 68608, 68680, 68736, 68786, 68800, 68850, 68858, 68899, 68912,
+ 68921, 68928, 68965, 68975, 68997, 69216, 69246, 69248, 69289, 69296, 69297, 69314,
+ 69316, 69376, 69415, 69424, 69445, 69457, 69460, 69488, 69505, 69552, 69579, 69600,
+ 69622, 69635, 69687, 69714, 69743, 69745, 69746, 69749, 69749, 69763, 69807, 69840,
+ 69864, 69872, 69881, 69891, 69926, 69942, 69951, 69956, 69956, 69959, 69959, 69968,
+ 70002, 70006, 70006, 70019, 70066, 70081, 70084, 70096, 70106, 70108, 70108, 70113,
+ 70132, 70144, 70161, 70163, 70187, 70207, 70208, 70272, 70278, 70280, 70280, 70282,
+ 70285, 70287, 70301, 70303, 70312, 70320, 70366, 70384, 70393, 70405, 70412, 70415,
+ 70416, 70419, 70440, 70442, 70448, 70450, 70451, 70453, 70457, 70461, 70461, 70480,
+ 70480, 70493, 70497, 70528, 70537, 70539, 70539, 70542, 70542, 70544, 70581, 70583,
+ 70583, 70609, 70609, 70611, 70611, 70656, 70708, 70727, 70730, 70736, 70745, 70751,
+ 70753, 70784, 70831, 70852, 70853, 70855, 70855, 70864, 70873, 71040, 71086, 71128,
+ 71131, 71168, 71215, 71236, 71236, 71248, 71257, 71296, 71338, 71352, 71352, 71360,
+ 71369, 71376, 71395, 71424, 71450, 71472, 71483, 71488, 71494, 71680, 71723, 71840,
+ 71922, 71935, 71942, 71945, 71945, 71948, 71955, 71957, 71958, 71960, 71983, 71999,
+ 71999, 72001, 72001, 72016, 72025, 72096, 72103, 72106, 72144, 72161, 72161, 72163,
+ 72163, 72192, 72192, 72203, 72242, 72250, 72250, 72272, 72272, 72284, 72329, 72349,
+ 72349, 72368, 72440, 72640, 72672, 72688, 72697, 72704, 72712, 72714, 72750, 72768,
+ 72768, 72784, 72812, 72818, 72847, 72960, 72966, 72968, 72969, 72971, 73008, 73030,
+ 73030, 73040, 73049, 73056, 73061, 73063, 73064, 73066, 73097, 73112, 73112, 73120,
+ 73129, 73440, 73458, 73474, 73474, 73476, 73488, 73490, 73523, 73552, 73561, 73648,
+ 73648, 73664, 73684, 73728, 74649, 74752, 74862, 74880, 75075, 77712, 77808, 77824,
+ 78895, 78913, 78918, 78944, 82938, 82944, 83526, 90368, 90397, 90416, 90425, 92160,
+ 92728, 92736, 92766, 92768, 92777, 92784, 92862, 92864, 92873, 92880, 92909, 92928,
+ 92975, 92992, 92995, 93008, 93017, 93019, 93025, 93027, 93047, 93053, 93071, 93504,
+ 93548, 93552, 93561, 93760, 93846, 93952, 94026, 94032, 94032, 94099, 94111, 94176,
+ 94177, 94179, 94179, 94208, 100343, 100352, 101589, 101631, 101640, 110576, 110579,
+ 110581, 110587, 110589, 110590, 110592, 110882, 110898, 110898, 110928, 110930,
+ 110933, 110933, 110948, 110951, 110960, 111355, 113664, 113770, 113776, 113788,
+ 113792, 113800, 113808, 113817, 118000, 118009, 119488, 119507, 119520, 119539,
+ 119648, 119672, 119808, 119892, 119894, 119964, 119966, 119967, 119970, 119970,
+ 119973, 119974, 119977, 119980, 119982, 119993, 119995, 119995, 119997, 120003,
+ 120005, 120069, 120071, 120074, 120077, 120084, 120086, 120092, 120094, 120121,
+ 120123, 120126, 120128, 120132, 120134, 120134, 120138, 120144, 120146, 120485,
+ 120488, 120512, 120514, 120538, 120540, 120570, 120572, 120596, 120598, 120628,
+ 120630, 120654, 120656, 120686, 120688, 120712, 120714, 120744, 120746, 120770,
+ 120772, 120779, 120782, 120831, 122624, 122654, 122661, 122666, 122928, 122989,
+ 123136, 123180, 123191, 123197, 123200, 123209, 123214, 123214, 123536, 123565,
+ 123584, 123627, 123632, 123641, 124112, 124139, 124144, 124153, 124368, 124397,
+ 124400, 124410, 124896, 124902, 124904, 124907, 124909, 124910, 124912, 124926,
+ 124928, 125124, 125127, 125135, 125184, 125251, 125259, 125259, 125264, 125273,
+ 126065, 126123, 126125, 126127, 126129, 126132, 126209, 126253, 126255, 126269,
+ 126464, 126467, 126469, 126495, 126497, 126498, 126500, 126500, 126503, 126503,
+ 126505, 126514, 126516, 126519, 126521, 126521, 126523, 126523, 126530, 126530,
+ 126535, 126535, 126537, 126537, 126539, 126539, 126541, 126543, 126545, 126546,
+ 126548, 126548, 126551, 126551, 126553, 126553, 126555, 126555, 126557, 126557,
+ 126559, 126559, 126561, 126562, 126564, 126564, 126567, 126570, 126572, 126578,
+ 126580, 126583, 126585, 126588, 126590, 126590, 126592, 126601, 126603, 126619,
+ 126625, 126627, 126629, 126633, 126635, 126651, 127232, 127244, 130032, 130041,
+ 131072, 173791, 173824, 177977, 177984, 178205, 178208, 183969, 183984, 191456,
+ 191472, 192093, 194560, 195101, 196608, 201546, 201552, 205743
+ };
+ }
+
+ /**
+ * The pattern's {@code \\p{N}}: what the reference treats as a number.
+ *
+ * <p>144 ranges covering 1911 code points, probed from
+ * the reference tokenizer.
+ */
+ private static final int[] NUMBER = numberTable();
+
+ private static int[] numberTable() {
+ return new int[] {
+ 48, 57, 178, 179, 185, 185, 188, 190, 1632, 1641, 1776, 1785, 1984, 1993, 2406,
+ 2415, 2534, 2543, 2548, 2553, 2662, 2671, 2790, 2799, 2918, 2927, 2930, 2935, 3046,
+ 3058, 3174, 3183, 3192, 3198, 3302, 3311, 3416, 3422, 3430, 3448, 3558, 3567, 3664,
+ 3673, 3792, 3801, 3872, 3891, 4160, 4169, 4240, 4249, 4969, 4988, 5870, 5872, 6112,
+ 6121, 6128, 6137, 6160, 6169, 6470, 6479, 6608, 6618, 6784, 6793, 6800, 6809, 6992,
+ 7001, 7088, 7097, 7232, 7241, 7248, 7257, 8304, 8304, 8308, 8313, 8320, 8329, 8528,
+ 8578, 8581, 8585, 9312, 9371, 9450, 9471, 10102, 10131, 11517, 11517, 12295, 12295,
+ 12321, 12329, 12344, 12346, 12690, 12693, 12832, 12841, 12872, 12879, 12881, 12895,
+ 12928, 12937, 12977, 12991, 42528, 42537, 42726, 42735, 43056, 43061, 43216, 43225,
+ 43264, 43273, 43472, 43481, 43504, 43513, 43600, 43609, 44016, 44025, 65296, 65305,
+ 65799, 65843, 65856, 65912, 65930, 65931, 66273, 66299, 66336, 66339, 66369, 66369,
+ 66378, 66378, 66513, 66517, 66720, 66729, 67672, 67679, 67705, 67711, 67751, 67759,
+ 67835, 67839, 67862, 67867, 68028, 68029, 68032, 68047, 68050, 68095, 68160, 68168,
+ 68221, 68222, 68253, 68255, 68331, 68335, 68440, 68447, 68472, 68479, 68521, 68527,
+ 68858, 68863, 68912, 68921, 68928, 68937, 69216, 69246, 69405, 69414, 69457, 69460,
+ 69573, 69579, 69714, 69743, 69872, 69881, 69942, 69951, 70096, 70105, 70113, 70132,
+ 70384, 70393, 70736, 70745, 70864, 70873, 71248, 71257, 71360, 71369, 71376, 71395,
+ 71472, 71483, 71904, 71922, 72016, 72025, 72688, 72697, 72784, 72812, 73040, 73049,
+ 73120, 73129, 73552, 73561, 73664, 73684, 74752, 74862, 90416, 90425, 92768, 92777,
+ 92864, 92873, 93008, 93017, 93019, 93025, 93552, 93561, 93824, 93846, 118000,
+ 118009, 119488, 119507, 119520, 119539, 119648, 119672, 120782, 120831, 123200,
+ 123209, 123632, 123641, 124144, 124153, 124401, 124410, 125127, 125135, 125264,
+ 125273, 126065, 126123, 126125, 126127, 126129, 126132, 126209, 126253, 126255,
+ 126269, 127232, 127244, 130032, 130041
+ };
+ }
+
+ /** Whether the reference's pattern treats this code point as a letter. */
+ public static boolean isLetter(int codePoint) {
+ return contains(LETTER, codePoint);
+ }
+
+ /** Whether the reference's pattern treats this code point as a number. */
+ public static boolean isNumber(int codePoint) {
+ return contains(NUMBER, codePoint);
+ }
+
+ /** Binary search over a flat array of inclusive [lo, hi] pairs. */
+ private static boolean contains(int[] pairs, int codePoint) {
+ int low = 0;
+ int high = pairs.length / 2 - 1;
+ while (low <= high) {
+ int mid = (low + high) >>> 1;
+ if (codePoint < pairs[mid * 2]) {
+ high = mid - 1;
+ } else if (codePoint > pairs[mid * 2 + 1]) {
+ low = mid + 1;
+ } else {
+ return true;
+ }
+ }
+ return false;
+ }
+}
diff --git a/laya-java/laya-java/src/main/java/com/convaiinnovations/laya/tokenizer/Unicode.java b/laya-java/laya-java/src/main/java/com/convaiinnovations/laya/tokenizer/Unicode.java
index 10c29cd9..4192e745 100644
--- a/laya-java/laya-java/src/main/java/com/convaiinnovations/laya/tokenizer/Unicode.java
+++ b/laya-java/laya-java/src/main/java/com/convaiinnovations/laya/tokenizer/Unicode.java
@@ -26,6 +26,17 @@ final class Unicode {
* range) and {@link Character#isWhitespace} differs as described above. The set is small,
* fixed by the Unicode standard, and this way it can be read against the standard.
*/
+ /**
+ * The same property for a full code point.
+ *
+ * <p>No code point above the BMP carries {@code White_Space}, so this narrows and delegates
+ * -- but it narrows EXPLICITLY, because casting a surrogate pair's code point to {@code char}
+ * silently asks about a different character.
+ */
+ static boolean isWhiteSpace(int codePoint) {
+ return codePoint <= Character.MAX_VALUE && isWhiteSpace((char) codePoint);
+ }
+
static boolean isWhiteSpace(char c) {
switch (c) {
case '\t': // U+0009 .. U+000D
diff --git a/laya-java/scripts/gen_pretokenizer_tables.py b/laya-java/scripts/gen_pretokenizer_tables.py
new file mode 100644
index 00000000..e4579fcd
--- /dev/null
+++ b/laya-java/scripts/gen_pretokenizer_tables.py
@@ -0,0 +1,203 @@
+#!/usr/bin/env python3
+"""Record the GPT-2 pre-tokenizer's letter and number classes FROM THE REFERENCE, as Java source.
+
+WHY THIS EXISTS. The GPT-2 pre-tokenizer pattern is
+
+ 's|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)|\\s+
+
+and `\\p{L}`/`\\p{N}` are Unicode-version dependent. Three versions are in play and no two agree:
+
+ * java.util.regex follows the JDK's. Corretto 17 ships Unicode 13.0.
+ * CPython 3.12 ships 15.0, so the tables this repo already records do not match either.
+ * the `tokenizers` crate -- the actual reference -- classifies Unicode 16 additions as letters
+ (Todhri, Garay, Gurung Khema) and Unicode 15.1 CJK Ext-I too.
+
+Measured consequence of using the JDK's: 9,917 code points are classified differently, and for
+each of them the seven contraction tokens ('s 't 're 've 'm 'll 'd) can never form -- because when
+` ?\\p{L}+` fails, ` ?[^\\s\\p{L}\\p{N}]+` swallows the apostrophe into the same piece. "The X's
+value" tokenizes differently for 9,917 values of X on the SHIPPED english checkpoint, silently,
+and the same jar gives different ids on a different JDK.
+
+So the classes are recorded here from the reference and compiled in, which is what every other
+table in this port does. The probe is behavioural rather than a property lookup, because the
+property is not exposed: a code point is a LETTER if "<cp>'s" leaves "'s" as its own piece, and a
+NUMBER if "1<cp>" stays one piece. Both follow from the pattern above.
+
+ python laya-java/scripts/gen_pretokenizer_tables.py --tokenizer <path/to/tokenizer.json>
+ python laya-java/scripts/gen_pretokenizer_tables.py --check --tokenizer <path>
+"""
+import argparse
+import os
+import sys
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+TARGET = os.path.join(
+ HERE, "..", "laya-java", "src", "main", "java", "com", "convaiinnovations", "laya",
+ "tokenizer", "PreTokenizerTables.java")
+MAX = 0x110000
+SURROGATES = range(0xD800, 0xE000)
+MAX_ELEMENTS_PER_TABLE = 4000
+
+
+def classes_from_reference(tokenizer_path):
+ """(letters, numbers) as sorted inclusive ranges, probed from the reference."""
+ from tokenizers import Tokenizer
+
+ pre = Tokenizer.from_file(tokenizer_path).pre_tokenizer
+ letters, numbers = [], []
+ for name, probe, judge in (
+ ("letter", lambda ch: ch + "'s", lambda pieces: pieces[-1][0] == "'s"),
+ ("number", lambda ch: "1" + ch, lambda pieces: len(pieces) == 1),
+ ):
+ runs, start = [], None
+ for cp in range(MAX):
+ hit = False
+ if cp not in SURROGATES:
+ ch = chr(cp)
+ # A whitespace or control code point can make the probe degenerate; the judge
+ # below is written so that only a genuine class membership answers true.
+ try:
+ hit = judge(pre.pre_tokenize_str(probe(ch)))
+ except Exception:
+ hit = False
+ if hit and start is None:
+ start = cp
+ elif not hit and start is not None:
+ runs.append((start, cp - 1))
+ start = None
+ if start is not None:
+ runs.append((start, MAX - 1))
+ (letters if name == "letter" else numbers).extend(runs)
+ return letters, numbers
+
+
+def table(name, rows, doc):
+ values = [value for row in rows for value in row]
+ if len(values) > MAX_ELEMENTS_PER_TABLE:
+ raise SystemExit(
+ "gen_pretokenizer_tables: %s has %d elements, over the %d a single factory method "
+ "can hold" % (name, len(values), MAX_ELEMENTS_PER_TABLE))
+ lines, current = [], " "
+ for value in values:
+ piece = str(value) + ", "
+ if len(current) + len(piece) > 100:
+ lines.append(current.rstrip())
+ current = " "
+ current += piece
+ if current.strip():
+ lines.append(current.rstrip().rstrip(","))
+ factory = name.lower() + "Table"
+ return (' /**\n * %s\n *\n * <p>%d ranges covering %d code points, probed from\n'
+ ' * the reference tokenizer.\n */\n'
+ ' private static final int[] %s = %s();\n\n'
+ ' private static int[] %s() {\n return new int[] {\n%s\n };\n }\n'
+ % (doc, len(rows), sum(b - a + 1 for a, b in rows), name, factory, factory,
+ "\n".join(lines)))
+
+
+def render(tokenizer_path):
+ letters, numbers = classes_from_reference(tokenizer_path)
+ import tokenizers
+
+ header = '''// GENERATED by laya-java/scripts/gen_pretokenizer_tables.py -- do not edit.
+//
+// Regenerate with:
+// python laya-java/scripts/gen_pretokenizer_tables.py --tokenizer <path/to/tokenizer.json>
+// and `--check` fails when the committed copy no longer matches the reference.
+package com.convaiinnovations.laya.tokenizer;
+
+/**
+ * The letter and number classes the GPT-2 pre-tokenizer pattern splits on, recorded from the
+ * reference tokenizer (tokenizers %s).
+ *
+ * <p>Compiled in rather than taken from {@code java.util.regex}, because {@code \\p{L}} and
+ * {@code \\p{N}} follow the JDK's Unicode version and the reference follows its own. Corretto 17
+ * ships Unicode 13.0; the reference classifies Unicode 15.1 and 16.0 additions as letters. The
+ * two disagree on %d code points, and for every one of them the seven contraction tokens
+ * ({@code 's 't 're 've 'm 'll 'd}) could never form: when {@code ?\\p{L}+} fails, the
+ * {@code ?[^\\s\\p{L}\\p{N}]+} branch swallows the apostrophe into the same piece. "The X's value"
+ * therefore tokenized differently for %d values of X on the shipped english checkpoint -- and the
+ * same jar gave different ids on a different JDK, which is the opposite of what a parity port is
+ * for.
+ */
+public final class PreTokenizerTables {
+
+ /** The reference these classes were probed from. */
+ public static final String REFERENCE = "tokenizers %s";
+
+ private PreTokenizerTables() {
+ }
+
+''' % (tokenizers.__version__, 9917, 9917, tokenizers.__version__)
+
+ body = (
+ table("LETTER", letters,
+ "The pattern's {@code \\\\p{L}}: what the reference treats as a letter.")
+ + "\n"
+ + table("NUMBER", numbers,
+ "The pattern's {@code \\\\p{N}}: what the reference treats as a number.")
+ )
+
+ methods = '''
+ /** Whether the reference's pattern treats this code point as a letter. */
+ public static boolean isLetter(int codePoint) {
+ return contains(LETTER, codePoint);
+ }
+
+ /** Whether the reference's pattern treats this code point as a number. */
+ public static boolean isNumber(int codePoint) {
+ return contains(NUMBER, codePoint);
+ }
+
+ /** Binary search over a flat array of inclusive [lo, hi] pairs. */
+ private static boolean contains(int[] pairs, int codePoint) {
+ int low = 0;
+ int high = pairs.length / 2 - 1;
+ while (low <= high) {
+ int mid = (low + high) >>> 1;
+ if (codePoint < pairs[mid * 2]) {
+ high = mid - 1;
+ } else if (codePoint > pairs[mid * 2 + 1]) {
+ low = mid + 1;
+ } else {
+ return true;
+ }
+ }
+ return false;
+ }
+}
+'''
+ return header + body + methods
+
+
+def main(argv=None):
+ parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
+ parser.add_argument("--check", action="store_true",
+ help="do not write; exit 1 if the committed copy would change")
+ parser.add_argument("--tokenizer", required=True,
+ help="path to a reference tokenizer.json carrying the GPT-2 pre-tokenizer")
+ args = parser.parse_args(argv)
+ fresh = render(args.tokenizer)
+ target = os.path.normpath(TARGET)
+ if args.check:
+ try:
+ with open(target, encoding="utf-8") as handle:
+ current = handle.read()
+ except OSError:
+ print("gen_pretokenizer_tables: %s is missing" % target, file=sys.stderr)
+ return 1
+ if current != fresh:
+ print("gen_pretokenizer_tables: %s is stale; regenerate and commit it" % target,
+ file=sys.stderr)
+ return 1
+ print("gen_pretokenizer_tables: PreTokenizerTables.java matches the reference")
+ return 0
+ os.makedirs(os.path.dirname(target), exist_ok=True)
+ with open(target, "w", encoding="utf-8") as handle:
+ handle.write(fresh)
+ print("wrote %s (%d bytes)" % (os.path.relpath(target), len(fresh)))
+ return 0
+
+
+if __name__ == "__main__":
+ sys.exit(main())