kitoken 0.11.0

Fast tokenizer for language models, supporting BPE, Unigram and WordPiece tokenization
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270

#[cfg(feature = "std")]
use std::fs::File;
#[cfg(feature = "std")]
use std::io::Read;
#[cfg(feature = "std")]
use std::path::Path;

use alloc::format;
use alloc::string::ToString;

use crate::convert::ConversionError;
use crate::{
    Configuration, Definition, Fallback, InsertionPosition, Kitoken, Metadata, Model, Regex,
    SpecialToken, SpecialTokenKind, SpecialVocab, Split, SplitBehavior, Template, Token, Vocab,
};

mod ms {
    use alloc::string::{String, ToString};
    use alloc::vec::Vec;
    use base64::{Engine, alphabet, engine};
    use serde::{Deserialize, Deserializer};

    static BASE64: engine::GeneralPurpose =
        const { engine::GeneralPurpose::new(&alphabet::STANDARD, engine::general_purpose::PAD) };

    fn from_base64<'de, D>(deserializer: D) -> Result<Vec<u8>, D::Error>
    where
        D: Deserializer<'de>, {
        let s: &str = Deserialize::deserialize(deserializer)?;
        let precompiled_charsmap =
            BASE64.decode(s).map_err(|e| serde::de::Error::custom(e.to_string()))?.to_vec();
        Ok(precompiled_charsmap)
    }

    #[derive(Deserialize, Debug, Clone, PartialEq)]
    pub struct Config {
        pub pattern:                    String,
        #[allow(unused)]
        pub num_vocab_tokens:           Option<usize>,
        #[allow(unused)]
        pub default_vocab_size:         Option<usize>,
        pub default_num_special_tokens: Option<usize>,
        pub version:                    String,
    }

    #[derive(Deserialize, Debug, Clone, PartialEq)]
    pub struct Token {
        pub rank:        usize,
        #[serde(deserialize_with = "from_base64")]
        pub token_bytes: Vec<u8>,
        #[allow(unused)]
        pub token_str:   Option<String>,
    }

    #[derive(Deserialize, Debug, Clone, PartialEq)]
    pub struct Tokenizer {
        pub config: Config,
        pub vocab:  Vec<Token>,
    }
}

use ms::Tokenizer;

/// Converts a `tekken` tokenizer definition into the definition format used by this crate.
///
/// `data` is the JSON data used by the `tekken` library, commonly stored as `tekken.json`.
///
/// Returns the tokenizer definition, or an error if the conversion fails.
///
/// # Examples
///
/// ```
/// # fn main() -> Result<(), Box<dyn std::error::Error>> {
/// use kitoken::Kitoken;
/// use kitoken::convert::convert_tekken;
///
/// let data = std::fs::read("tests/models/tekken/mistral2410.json")?;
/// let definition = convert_tekken(data).unwrap();
///
/// let tokenizer = Kitoken::try_from(definition).unwrap();
/// # Ok(())
/// # }
/// ```
///
///  Additional conversion utilities are defined in [`Definition`] and [`Kitoken`].
///
/// # Format
///
/// The `tekken` definition is composed of a JSON object with the following fields:
///
/// - `config`: The model configuration with the following fields:
///   - `pattern`: The regex pattern used to split the input.
///   - `num_vocab_tokens`: The number of vocabulary tokens.
///   - `default_vocab_size`: The default vocabulary size.
///   - `default_num_special_tokens`: The default number of special tokens.
///   - `version`: The version of the model.
/// - `vocab`: The vocabulary with elements with the following fields:
///   - `rank`: The rank of the token.
///   - `token_bytes`: The token bytes.
///   - `token_str`: The string representation of the token.
///
/// See the [tekken documentation](https://docs.mistral.ai/guides/tokenization/) for more information.
pub fn convert_tekken(data: impl AsRef<[u8]>) -> Result<Definition, ConversionError> {
    let data = data.as_ref();

    let tokenizer = serde_json::from_slice::<Tokenizer>(data)
        .map_err(|e| ConversionError::InvalidData(format!("invalid JSON: {}", e)))?;

    let specials = &[
        ("<unk>", Some("unk".to_string()), false),
        ("<s>", Some("bos".to_string()), false),
        ("</s>", Some("eos".to_string()), false),
        ("[INST]", None, true),
        ("[/INST]", None, true),
        ("[AVAILABLE_TOOLS]", None, true),
        ("[/AVAILABLE_TOOLS]", None, true),
        ("[TOOL_RESULTS]", None, true),
        ("[/TOOL_RESULTS]", None, true),
        ("[TOOL_CALLS]", None, true),
        ("<pad>", Some("pad".to_string()), false),
        ("[PREFIX]", None, true),
        ("[MIDDLE]", None, true),
        ("[SUFFIX]", None, true),
    ];

    let specials_len = tokenizer.config.default_num_special_tokens.unwrap_or(specials.len());
    let vocab_len = tokenizer.config.default_vocab_size.unwrap_or(tokenizer.vocab.len());
    if vocab_len > tokenizer.vocab.len() + specials_len {
        return Err(ConversionError::InvalidData(format!(
            "too many tokens: {} > {}",
            vocab_len,
            tokenizer.vocab.len() + specials_len
        )));
    }

    if vocab_len >= u32::MAX as usize {
        return Err(ConversionError::InvalidData(format!(
            "too many pieces in vocab: {}",
            vocab_len
        )));
    }
    if specials_len >= u32::MAX as usize {
        return Err(ConversionError::InvalidData(format!(
            "too many pieces in specials: {}",
            specials_len
        )));
    }

    let mut config = Configuration::default();
    config.fallback.push(Fallback::Unknown);
    config.fallback.push(Fallback::Skip);

    config.split.push(Split::Pattern {
        pattern:  Regex::new(&tokenizer.config.pattern)?.into(),
        behavior: SplitBehavior::Isolate,
    });

    let mut specials = specials
        .iter()
        .enumerate()
        .map(|(i, (s, d, e))| SpecialToken {
            id:      i as u32,
            bytes:   s.as_bytes().to_vec(),
            kind:    SpecialTokenKind::Control,
            ident:   d.clone(),
            score:   i as f32,
            extract: *e,
        })
        .collect::<SpecialVocab>();
    specials[0].kind = SpecialTokenKind::Unknown;
    if specials.len() < specials_len {
        for i in specials.len()..specials_len {
            specials.push(SpecialToken {
                id:      i as u32,
                bytes:   format!("<SPECIAL_{}>", i).as_bytes().to_vec(),
                kind:    SpecialTokenKind::Control,
                ident:   None,
                score:   i as f32,
                extract: true,
            });
        }
    }
    specials.sort();

    let mut vocab = Vocab::with_capacity(vocab_len);
    // This will throw away any tokens beyond the set vocab size.
    // This is consistent with the behavior of `tekken`.
    for token in tokenizer.vocab.into_iter().take(vocab_len - specials.len()) {
        vocab.push(Token {
            id:    token.rank as u32 + specials.len() as u32,
            bytes: token.token_bytes,
        });
    }
    vocab.sort();

    let model = Model::BytePair {
        vocab,
        chars: false,
    };

    config.templates.push(Template {
        content:  "<s>".to_string(),
        position: InsertionPosition::SequenceStart,
    });
    config.templates.push(Template {
        content:  "</s>".to_string(),
        position: InsertionPosition::SequenceEnd,
    });

    let meta = Metadata {
        source: "tekken".to_string(),
        ..Metadata::default()
    };

    Ok(Definition {
        meta,
        model,
        specials,
        config,
    })
}


impl Definition {
    /// Converts a `tekken` model into the encoder format used by this crate.
    /// See [`convert_tekken`] for more details.
    #[cfg(feature = "std")]
    pub fn from_tekken_reader<R: Read>(reader: &mut R) -> Result<Self, ConversionError> {
        use alloc::vec::Vec;
        let mut data = Vec::with_capacity(1024);
        reader.read_to_end(&mut data)?;
        Self::from_tekken_slice(&data)
    }

    /// Converts a `tekken` model into the encoder format used by this crate.
    /// See [`convert_tekken`] for more details.
    #[cfg(feature = "std")]
    pub fn from_tekken_file(path: impl AsRef<Path>) -> Result<Self, ConversionError> {
        let mut file = File::open(path)?;
        Self::from_tekken_reader(&mut file)
    }

    /// Converts a `tekken` tokenizer definition into the encoder format used by this crate.
    /// See [`convert_tekken`] for more details.
    pub fn from_tekken_slice(data: &[u8]) -> Result<Self, ConversionError> {
        convert_tekken(data)
    }
}

impl Kitoken {
    /// Initializes the tokenizer from a `tekken` model.
    /// See [`convert_tekken`] for more details.
    #[cfg(feature = "std")]
    pub fn from_tekken_reader<R: Read>(reader: &mut R) -> Result<Self, ConversionError> {
        Ok(Self::from_definition(Definition::from_tekken_reader(reader)?)?)
    }

    /// Initializes the tokenizer from a `tekken` model.
    /// See [`convert_tekken`] for more details.
    #[cfg(feature = "std")]
    pub fn from_tekken_file(path: impl AsRef<Path>) -> Result<Self, ConversionError> {
        Ok(Self::from_definition(Definition::from_tekken_file(path)?)?)
    }

    /// Initializes the tokenizer from a `tekken` model.
    /// See [`convert_tekken`] for more details.
    pub fn from_tekken_slice(data: &[u8]) -> Result<Self, ConversionError> {
        Ok(Self::from_definition(Definition::from_tekken_slice(data)?)?)
    }
}