Add tokenizer

LICENSE

@@ -0,0 +1,339 @@
+                    GNU GENERAL PUBLIC LICENSE
+                       Version 2, June 1991
+
+ Copyright (C) 1989, 1991 Free Software Foundation, Inc.,
+ 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The licenses for most software are designed to take away your
+freedom to share and change it.  By contrast, the GNU General Public
+License is intended to guarantee your freedom to share and change free
+software--to make sure the software is free for all its users.  This
+General Public License applies to most of the Free Software
+Foundation's software and to any other program whose authors commit to
+using it.  (Some other Free Software Foundation software is covered by
+the GNU Lesser General Public License instead.)  You can apply it to
+your programs, too.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+this service if you wish), that you receive source code or can get it
+if you want it, that you can change the software or use pieces of it
+in new free programs; and that you know you can do these things.
+
+  To protect your rights, we need to make restrictions that forbid
+anyone to deny you these rights or to ask you to surrender the rights.
+These restrictions translate to certain responsibilities for you if you
+distribute copies of the software, or if you modify it.
+
+  For example, if you distribute copies of such a program, whether
+gratis or for a fee, you must give the recipients all the rights that
+you have.  You must make sure that they, too, receive or can get the
+source code.  And you must show them these terms so they know their
+rights.
+
+  We protect your rights with two steps: (1) copyright the software, and
+(2) offer you this license which gives you legal permission to copy,
+distribute and/or modify the software.
+
+  Also, for each author's protection and ours, we want to make certain
+that everyone understands that there is no warranty for this free
+software.  If the software is modified by someone else and passed on, we
+want its recipients to know that what they have is not the original, so
+that any problems introduced by others will not reflect on the original
+authors' reputations.
+
+  Finally, any free program is threatened constantly by software
+patents.  We wish to avoid the danger that redistributors of a free
+program will individually obtain patent licenses, in effect making the
+program proprietary.  To prevent this, we have made it clear that any
+patent must be licensed for everyone's free use or not licensed at all.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                    GNU GENERAL PUBLIC LICENSE
+   TERMS AND CONDITIONS FOR COPYING, DISTRIBUTION AND MODIFICATION
+
+  0. This License applies to any program or other work which contains
+a notice placed by the copyright holder saying it may be distributed
+under the terms of this General Public License.  The "Program", below,
+refers to any such program or work, and a "work based on the Program"
+means either the Program or any derivative work under copyright law:
+that is to say, a work containing the Program or a portion of it,
+either verbatim or with modifications and/or translated into another
+language.  (Hereinafter, translation is included without limitation in
+the term "modification".)  Each licensee is addressed as "you".
+
+Activities other than copying, distribution and modification are not
+covered by this License; they are outside its scope.  The act of
+running the Program is not restricted, and the output from the Program
+is covered only if its contents constitute a work based on the
+Program (independent of having been made by running the Program).
+Whether that is true depends on what the Program does.
+
+  1. You may copy and distribute verbatim copies of the Program's
+source code as you receive it, in any medium, provided that you
+conspicuously and appropriately publish on each copy an appropriate
+copyright notice and disclaimer of warranty; keep intact all the
+notices that refer to this License and to the absence of any warranty;
+and give any other recipients of the Program a copy of this License
+along with the Program.
+
+You may charge a fee for the physical act of transferring a copy, and
+you may at your option offer warranty protection in exchange for a fee.
+
+  2. You may modify your copy or copies of the Program or any portion
+of it, thus forming a work based on the Program, and copy and
+distribute such modifications or work under the terms of Section 1
+above, provided that you also meet all of these conditions:
+
+    a) You must cause the modified files to carry prominent notices
+    stating that you changed the files and the date of any change.
+
+    b) You must cause any work that you distribute or publish, that in
+    whole or in part contains or is derived from the Program or any
+    part thereof, to be licensed as a whole at no charge to all third
+    parties under the terms of this License.
+
+    c) If the modified program normally reads commands interactively
+    when run, you must cause it, when started running for such
+    interactive use in the most ordinary way, to print or display an
+    announcement including an appropriate copyright notice and a
+    notice that there is no warranty (or else, saying that you provide
+    a warranty) and that users may redistribute the program under
+    these conditions, and telling the user how to view a copy of this
+    License.  (Exception: if the Program itself is interactive but
+    does not normally print such an announcement, your work based on
+    the Program is not required to print an announcement.)
+
+These requirements apply to the modified work as a whole.  If
+identifiable sections of that work are not derived from the Program,
+and can be reasonably considered independent and separate works in
+themselves, then this License, and its terms, do not apply to those
+sections when you distribute them as separate works.  But when you
+distribute the same sections as part of a whole which is a work based
+on the Program, the distribution of the whole must be on the terms of
+this License, whose permissions for other licensees extend to the
+entire whole, and thus to each and every part regardless of who wrote it.
+
+Thus, it is not the intent of this section to claim rights or contest
+your rights to work written entirely by you; rather, the intent is to
+exercise the right to control the distribution of derivative or
+collective works based on the Program.
+
+In addition, mere aggregation of another work not based on the Program
+with the Program (or with a work based on the Program) on a volume of
+a storage or distribution medium does not bring the other work under
+the scope of this License.
+
+  3. You may copy and distribute the Program (or a work based on it,
+under Section 2) in object code or executable form under the terms of
+Sections 1 and 2 above provided that you also do one of the following:
+
+    a) Accompany it with the complete corresponding machine-readable
+    source code, which must be distributed under the terms of Sections
+    1 and 2 above on a medium customarily used for software interchange; or,
+
+    b) Accompany it with a written offer, valid for at least three
+    years, to give any third party, for a charge no more than your
+    cost of physically performing source distribution, a complete
+    machine-readable copy of the corresponding source code, to be
+    distributed under the terms of Sections 1 and 2 above on a medium
+    customarily used for software interchange; or,
+
+    c) Accompany it with the information you received as to the offer
+    to distribute corresponding source code.  (This alternative is
+    allowed only for noncommercial distribution and only if you
+    received the program in object code or executable form with such
+    an offer, in accord with Subsection b above.)
+
+The source code for a work means the preferred form of the work for
+making modifications to it.  For an executable work, complete source
+code means all the source code for all modules it contains, plus any
+associated interface definition files, plus the scripts used to
+control compilation and installation of the executable.  However, as a
+special exception, the source code distributed need not include
+anything that is normally distributed (in either source or binary
+form) with the major components (compiler, kernel, and so on) of the
+operating system on which the executable runs, unless that component
+itself accompanies the executable.
+
+If distribution of executable or object code is made by offering
+access to copy from a designated place, then offering equivalent
+access to copy the source code from the same place counts as
+distribution of the source code, even though third parties are not
+compelled to copy the source along with the object code.
+
+  4. You may not copy, modify, sublicense, or distribute the Program
+except as expressly provided under this License.  Any attempt
+otherwise to copy, modify, sublicense or distribute the Program is
+void, and will automatically terminate your rights under this License.
+However, parties who have received copies, or rights, from you under
+this License will not have their licenses terminated so long as such
+parties remain in full compliance.
+
+  5. You are not required to accept this License, since you have not
+signed it.  However, nothing else grants you permission to modify or
+distribute the Program or its derivative works.  These actions are
+prohibited by law if you do not accept this License.  Therefore, by
+modifying or distributing the Program (or any work based on the
+Program), you indicate your acceptance of this License to do so, and
+all its terms and conditions for copying, distributing or modifying
+the Program or works based on it.
+
+  6. Each time you redistribute the Program (or any work based on the
+Program), the recipient automatically receives a license from the
+original licensor to copy, distribute or modify the Program subject to
+these terms and conditions.  You may not impose any further
+restrictions on the recipients' exercise of the rights granted herein.
+You are not responsible for enforcing compliance by third parties to
+this License.
+
+  7. If, as a consequence of a court judgment or allegation of patent
+infringement or for any other reason (not limited to patent issues),
+conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot
+distribute so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you
+may not distribute the Program at all.  For example, if a patent
+license would not permit royalty-free redistribution of the Program by
+all those who receive copies directly or indirectly through you, then
+the only way you could satisfy both it and this License would be to
+refrain entirely from distribution of the Program.
+
+If any portion of this section is held invalid or unenforceable under
+any particular circumstance, the balance of the section is intended to
+apply and the section as a whole is intended to apply in other
+circumstances.
+
+It is not the purpose of this section to induce you to infringe any
+patents or other property right claims or to contest validity of any
+such claims; this section has the sole purpose of protecting the
+integrity of the free software distribution system, which is
+implemented by public license practices.  Many people have made
+generous contributions to the wide range of software distributed
+through that system in reliance on consistent application of that
+system; it is up to the author/donor to decide if he or she is willing
+to distribute software through any other system and a licensee cannot
+impose that choice.
+
+This section is intended to make thoroughly clear what is believed to
+be a consequence of the rest of this License.
+
+  8. If the distribution and/or use of the Program is restricted in
+certain countries either by patents or by copyrighted interfaces, the
+original copyright holder who places the Program under this License
+may add an explicit geographical distribution limitation excluding
+those countries, so that distribution is permitted only in or among
+countries not thus excluded.  In such case, this License incorporates
+the limitation as if written in the body of this License.
+
+  9. The Free Software Foundation may publish revised and/or new versions
+of the General Public License from time to time.  Such new versions will
+be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+Each version is given a distinguishing version number.  If the Program
+specifies a version number of this License which applies to it and "any
+later version", you have the option of following the terms and conditions
+either of that version or of any later version published by the Free
+Software Foundation.  If the Program does not specify a version number of
+this License, you may choose any version ever published by the Free Software
+Foundation.
+
+  10. If you wish to incorporate parts of the Program into other free
+programs whose distribution conditions are different, write to the author
+to ask for permission.  For software which is copyrighted by the Free
+Software Foundation, write to the Free Software Foundation; we sometimes
+make exceptions for this.  Our decision will be guided by the two goals
+of preserving the free status of all derivatives of our free software and
+of promoting the sharing and reuse of software generally.
+
+                            NO WARRANTY
+
+  11. BECAUSE THE PROGRAM IS LICENSED FREE OF CHARGE, THERE IS NO WARRANTY
+FOR THE PROGRAM, TO THE EXTENT PERMITTED BY APPLICABLE LAW.  EXCEPT WHEN
+OTHERWISE STATED IN WRITING THE COPYRIGHT HOLDERS AND/OR OTHER PARTIES
+PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESSED
+OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.  THE ENTIRE RISK AS
+TO THE QUALITY AND PERFORMANCE OF THE PROGRAM IS WITH YOU.  SHOULD THE
+PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF ALL NECESSARY SERVICING,
+REPAIR OR CORRECTION.
+
+  12. IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MAY MODIFY AND/OR
+REDISTRIBUTE THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES,
+INCLUDING ANY GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING
+OUT OF THE USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED
+TO LOSS OF DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY
+YOU OR THIRD PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER
+PROGRAMS), EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGES.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+convey the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+Also add information on how to contact you by electronic and paper mail.
+
+If the program is interactive, make it output a short notice like this
+when it starts in an interactive mode:
+
+    Gnomovision version 69, Copyright (C) year name of author
+    Gnomovision comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
+    This is free software, and you are welcome to redistribute it
+    under certain conditions; type `show c' for details.
+
+The hypothetical commands `show w' and `show c' should show the appropriate
+parts of the General Public License.  Of course, the commands you use may
+be called something other than `show w' and `show c'; they could even be
+mouse-clicks or menu items--whatever suits your program.
+
+You should also get your employer (if you work as a programmer) or your
+school, if any, to sign a "copyright disclaimer" for the program, if
+necessary.  Here is a sample; alter the names:
+
+  Yoyodyne, Inc., hereby disclaims all copyright interest in the program
+  `Gnomovision' (which makes passes at compilers) written by James Hacker.
+
+  <signature of Ty Coon>, 1 April 1989
+  Ty Coon, President of Vice
+
+This General Public License does not permit incorporating your program into
+proprietary programs.  If your program is a subroutine library, you may
+consider it more useful to permit linking proprietary applications with the
+library.  If this is what you want to do, use the GNU Lesser General
+Public License instead of this License.

README.md

@@ -1,3 +1,67 @@
----
-license: gpl-2.0
----
+# Tokenizer
+
+Finetune here to talk a bit about [NovelAI](https://novelai.net/)'s new tokenizer that I worked on. First a quick reminder. In most cases, our models don't see words as individual letters. Instead, text is broken down into tokens, which are words or word fragments. For example, the sentence “`The quick brown fox jumps over the goblin.`” would tokenize as “`The| quick| brown| fox| jumps| over| the| go|bl|in.`” in the Pile tokenizer used by GPT-NeoX 20B and Krake, with each | signifying a boundary between tokens.
+
+When deciding on a tokenizer for a model, there are various criteria to consider. The first and most obvious is the vocabulary size. It may be tempting to just set it very high to ensure that every word or even multiple words, as in the case of the tokenizer used by AI21's Jurassic models, gets its own distinct token. However, this has the drawback that the model will be less able to generalize. That means it will not be able to make use of meaningful patterns in how words are spelt, such as similarities between words ending in “-ize”. It will also be less robust against misspellings. At the same time, the vocabulary of the tokenizer should not be too small. Common words should have their own token. The same goes for Unicode characters that are likely to show up in tokenized texts, because otherwise they will have to be constructed by the model byte-by-byte, which is much harder for the model. A good trade-off with regards to vocabulary size is around 32000 tokens for a single language vocabulary. This also has the benefit of fitting easily within 16 bits, which makes handling tokenized data easier in many cases.
+
+The type of tokenizer is another important decision to make. Unigram tokenizers have been shown to produce much more meaningful tokenizations of words, while so far the predominantly used tokenizer type for large language models (LLM) is BPE (byte pair encoding). The most common implementation of BPE is probably the GPT2 one, but Google's sentencepiece implementation of BPE offers the not so slight advantage of natively being able to tokenize Unicode characters directly, without having to assemble them from bytes, which requires additional tokens representing partial Unicode code points to be added to the vocabulary, wasting some additional space. For example, “🙂” consists of four bytes “`F0 9F 99 82`”, so in traditional BPE, `F0` would first get merged with `9F` to make up `F09F`, which is then merged with `99` to make up `F09F99`, which is then merged with `82`, so two additional intermediate tokens would have to be added to the vocabulary. At the same time, sentencepiece also supports tokenizing arbitrary binary data using byte tokens.
+
+Finally, the compression ratio achieved by the tokenizer is important to consider. If a given text tokenizes into less tokens, this will allow the LLM to see more text at once, given the fixed size of context it can see at a maximum, which is important for users of the LLm. It will also influence how much text you need to achieve a certain amount of tokens if, say, you are trying to meet a certain amount of training data. If your tokenizer compresses text less efficiently, you may more easily achieve a dataset of a given size, but it stands to reason that a model trained on such a less efficiently tokenized dataset of a given size will learn less than one trained of on a same sized dataset that was tokenized with a tokenizer that achieves a higher compression ratio, because in effect, it will see less bits of actually information during training.
+
+With all these things in mind, we decided that we want our own tokenizer for the models we will train, that is better optimized for our use cases, such as storytelling.
+
+Tokenizers are trained on data, so we started by extracting small randomized subsets from the various distinct subsets of our model training dataset and used these to evaluate the available tokenizer training approaches. Both Huggingface's tokenizers library and Google's sentencepiece support training tokenizers of different types. A preliminary investigation showed that sentencepiece's trainer is more memory efficient, although a training dataset in the low double digit gibibytes still required a compute node with 1TB of RAM to run successfully. Due to this, we decided to use sentencepiece.
+
+We originally decided on a vocabulary size of 32000, but when training Genji V2, we found that modifying an existing tokenizer to support an additional language was not a pleasant experience. As it seems likely that we will want to do similar [language transfer learning](https://blog.novelai.net/data-efficient-language-transfer-with-gpt-j-45daedaaf35a) in the future, we have decided to have our tokenizer accommodate both English and Japanese from the start. For this reason, we decided to double the vocabulary size to 64000, which then was close to filling up the available token ID space of 16 bits, so we went all the way to a vocabulary size of 65535 tokens. During tokenizer training, I carefully balanced the training data in such a way that latin alphabet tokens of a length of at least 2 characters and Japanese language tokens take up approximately the same amount of token space. Bumping the vocabulary size up to 65535 also allows more Unicode character tokens such as emoji. For the Japanese part of tokenizer training data, we used our existing Genji training data and a comparatively smaller amount of Japanese Wikipedia.
+
+We have manually added tokens for certain multi-whitespace strings and have set up the tokenizer in such a way that numbers are tokenized digit by digit. Tokenizing numbers digit by digit may slightly reduce compression ratio in number heavy texts, but it will also allow the LLM to more effectively learn how to handle numeric values.
+
+Considering the possible benefits of Unigram tokenizers, we started out by training a Unigram tokenizer. This took multiple runs of rebalancing the dataset between languages and also between the different subsets of our main datasets to get the token distribution to look the way we want. Each Unigram training run took a few hours. For the sake of comparison, we also trained a BPE model, which again required multiple runs to rebalance the dataset. BPE runs ended up much slower, taking nearly a whole day.
+
+Both tokenizers were then evaluated on a held-out part of the dataset. The idea was that, if the compression ratios are similar or Unigram is only slightly worse, we would use the Unigram tokenizer to benefit from the more natural word segmentation. We found that the BPE tokenizer has a 25-29% higher compression ratio on the largest parts of our English language dataset. This unexpectedly large gap in performance led us to choose the BPE tokenizer over the Unigram one and also explains the continuing prevalence of BPE tokenizers for LLMs. We also compared the compression ratio of our tokenizer to the LLaMa tokenizer, which is a sentencepiece based BPE tokenizer with a 32000 token vocabulary. In comparison to the LLaMa tokenizer, we find our tokenizer to achieve a 7-19% higher compression ratio on the largest parts of our English language dataset.
+
+Finally, I would like to give some stats about token distribution. Our tokenizer contains 28586 tokens made up of latin alphabet characters with a minimum length of two. Tokens with a leading space are included in this. It contains 18123 Japanese tokens longer than a single character and 9626 tokens for Japanese and Chinese characters, which cannot be easily told apart for the sake of these stats due to the Unicode han unification. 9200 other tokens are included. This space is taken up mostly by Unicode characters such as emoji.
+
+For comparison, the LLaMa tokenizer contains 23964 tokens made up only of latin alphabet characters, no Japanese token longer than a single character, 836 Japanese characters and 7224 other tokens.
+
+## JavaScript implementation
+
+The JavaScript implementation used by the NovelAI frontend can be found [here](https://github.com/NovelAI/nai-js-tokenizer).
+
+## V2
+
+For [V2](https://huggingface.co/NovelAI/nerdstash-tokenizer-v2/), the original digit special tokens were replaced with english contractions. Digits will therefore be encoded using corresponding the byte tokens instead.
+
+# Example usage with transformers
+
+Since it seems to be the most up-to-date class for using sentencepiece tokenizers in transformers, this tokenizer uses the `LlamaTokenizer` class. Note that the `LlamaTokenizerFast` class is not supported. `AutoTokenizer` selects the fast version and is also not supported.
+
+```python
+from transformers import LlamaTokenizer
+
+tokenizer = LlamaTokenizer.from_pretrained("NovelAI/nerdstash-tokenizer-v2")
+
+print(tokenizer.encode("Hello, world!"))
+```
+
+## Example usage with sentencepiece
+
+```python
+import sentencepiece as spm
+
+s = spm.SentencePieceProcessor(model_file='tokenizer.model')
+
+text = "The quick brown fox jumps over the goblin."
+
+print("Text:", text)
+
+print("Token IDs:", s.encode(text))
+# Token IDs: [541, 1939, 6573, 22820, 22734, 712, 336, 34477, 49230]
+
+print("Readable tokens:", s.encode(text, out_type=str))
+# Readable tokens: ['The', '▁quick', '▁brown', '▁fox', '▁jumps', '▁over', '▁the', '▁goblin', '.']
+```
+
+## License
+
+The tokenizer is licensed under the GNU General Public License, version 2.

special_tokens_map.json

@@ -0,0 +1 @@
+{"eos_token": "<|endoftext|>", "unk_token": "<|unknown|>", "pad_token": "<|pad|>", "bos_token": "<|startoftext|>", "additional_special_tokens": ["'t've", "'s", "'t", "'ve", "'n", "'d", "'ll", "'re", "'m", "'clock", "\u2581\u2581", "\u2581\u2581\u2581\u2581", "\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581", "\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581", "\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581", "\u2003", "\u2002", "\u2042", "\u2500", "***", "----", "\u2581\"", "\u2581Author", "\u2581Title", "\u2581Tags", "\u2581Genre", "\u2581Style", "\u2581Knowledge", "\u2581Summary", "\u2581Rating", "Type", "Characters", "Glossary", "<|spmspace|>", "<|spmspace|><|spmspace|>", "<|spmspace|><|spmspace|><|spmspace|><|spmspace|>", "<|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|>", "<|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|>", "<|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|>", "<|mtvocab|>", "<|mtvenglish|>", "<|mtvjapanese|>", "<|mtsentence|>", "<|mtsjapanese|>", "<|mtsenglish|>", "<|mtsentenceend|>", "<|mtvocabend|>", "<|mtend|>", "<|mask|>", "<|masksingle|>", "<|maskshort|>", "<|maskmedium|>", "<|masklong|>", "<|maskparagraph|>", "<|maskend|>", "<|fill|>", "<|fillend|>", "<|rubycover|>", "<|rubystart|>", "<|rubyend|>", "<|reserved0|>", "<|reserved1|>", "<|reserved2|>", "<|reserved3|>", "<|reserved4|>", "<|reserved5|>", "<|reserved6|>", "<|reserved7|>", "<|reserved8|>", "<|reserved9|>", "<|reserved10|>"]}

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:005ad680b10f1abd406bdb0ca9c6a5d83fc1f6e0a855bdd1942c1ceab1fb47ab
+size 1033744

tokenizer_config.json

@@ -0,0 +1 @@
+{"eos_token": "<|endoftext|>", "unk_token": "<|unknown|>", "pad_token": "<|pad|>", "bos_token": "<|startoftext|>", "additional_special_tokens": ["'t've", "'s", "'t", "'ve", "'n", "'d", "'ll", "'re", "'m", "'clock", "\u2581\u2581", "\u2581\u2581\u2581\u2581", "\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581", "\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581", "\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581\u2581", "\u2003", "\u2002", "\u2042", "\u2500", "***", "----", "\u2581\"", "\u2581Author", "\u2581Title", "\u2581Tags", "\u2581Genre", "\u2581Style", "\u2581Knowledge", "\u2581Summary", "\u2581Rating", "Type", "Characters", "Glossary", "<|spmspace|>", "<|spmspace|><|spmspace|>", "<|spmspace|><|spmspace|><|spmspace|><|spmspace|>", "<|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|>", "<|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|>", "<|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|><|spmspace|>", "<|mtvocab|>", "<|mtvenglish|>", "<|mtvjapanese|>", "<|mtsentence|>", "<|mtsjapanese|>", "<|mtsenglish|>", "<|mtsentenceend|>", "<|mtvocabend|>", "<|mtend|>", "<|mask|>", "<|masksingle|>", "<|maskshort|>", "<|maskmedium|>", "<|masklong|>", "<|maskparagraph|>", "<|maskend|>", "<|fill|>", "<|fillend|>", "<|rubycover|>", "<|rubystart|>", "<|rubyend|>", "<|reserved0|>", "<|reserved1|>", "<|reserved2|>", "<|reserved3|>", "<|reserved4|>", "<|reserved5|>", "<|reserved6|>", "<|reserved7|>", "<|reserved8|>", "<|reserved9|>", "<|reserved10|>"], "truncation": false, "model_max_length": 8192, "add_bos_token": true, "add_eos_token": false, "tokenizer_class": "LlamaTokenizer", "clean_up_tokenization_spaces": false, "sp_model_kwargs": {}}