tokie 0.0.10

Blazingly fast tokenizer - 50x faster tokenization, 10x smaller model files, 100% accurate drop-in replacement for HuggingFace
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

//! Token-accuracy tests: compare tokie against HuggingFace tokenizers on enwik8.
//!
//! Run with: cargo test -p tokie --test accuracy --features hf -- --ignored
//!
//! Requires network access and benches/data/enwik8 (1MB used).

#![cfg(feature = "hf")]

use std::path::Path;

use tokenizers::Tokenizer as HfTokenizer;
use tokie::Tokenizer;

/// Load first `max_bytes` of enwik8, returning valid UTF-8.
fn load_enwik8(max_bytes: usize) -> String {
    let path = Path::new(env!("CARGO_MANIFEST_DIR"))
        .parent()
        .unwrap()
        .parent()
        .unwrap()
        .join("benches/data/enwik8");
    let data = std::fs::read(&path).unwrap_or_else(|e| panic!("Failed to read {}: {e}", path.display()));
    let truncated = &data[..data.len().min(max_bytes)];
    String::from_utf8_lossy(truncated).into_owned()
}

/// Compare tokie (from tokiers/) against HuggingFace on enwik8.
/// Returns (pass, first_diff_index) — None means all tokens match.
fn compare_model(tokiers_repo: &str, hf_model: &str, text: &str) -> (bool, Option<usize>) {
    let tok = Tokenizer::from_pretrained(tokiers_repo)
        .unwrap_or_else(|e| panic!("Failed to load tokie {tokiers_repo}: {e}"));
    let mut hf = HfTokenizer::from_pretrained(hf_model, None)
        .unwrap_or_else(|e| panic!("Failed to load HF {hf_model}: {e}"));
    // Disable any default truncation so we compare the full output
    let _ = hf.with_truncation(None);

    let tokie_ids = tok.encode(text, false).ids;
    let hf_enc = hf.encode(text, false)
        .unwrap_or_else(|e| panic!("HF encode failed for {hf_model}: {e}"));
    let hf_ids = hf_enc.get_ids();

    if tokie_ids.as_slice() == hf_ids {
        (true, None)
    } else {
        let diff = tokie_ids.iter().zip(hf_ids.iter())
            .position(|(a, b)| a != b)
            .unwrap_or(tokie_ids.len().min(hf_ids.len()));
        (false, Some(diff))
    }
}

// ============================================================================
// WordPiece models (BERT-family)
// ============================================================================

macro_rules! accuracy_test {
    ($name:ident, $tokiers:expr, $hf:expr) => {
        #[test]
        #[ignore] // Requires network + enwik8
        fn $name() {
            let text = load_enwik8(1_000_000);
            let (pass, diff) = compare_model($tokiers, $hf, &text);
            assert!(pass, "Token mismatch at index {:?}", diff);
        }
    };
}

// WordPiece
accuracy_test!(bert_base_uncased,       "tokiers/bert-base-uncased",       "google-bert/bert-base-uncased");
accuracy_test!(all_minilm_l6_v2,        "tokiers/all-MiniLM-L6-v2",       "sentence-transformers/all-MiniLM-L6-v2");
accuracy_test!(all_minilm_l12_v2,       "tokiers/all-MiniLM-L12-v2",      "sentence-transformers/all-MiniLM-L12-v2");
accuracy_test!(all_mpnet_base_v2,       "tokiers/all-mpnet-base-v2",      "sentence-transformers/all-mpnet-base-v2");
accuracy_test!(bge_base_en_v1_5,        "tokiers/bge-base-en-v1.5",       "BAAI/bge-base-en-v1.5");
accuracy_test!(bge_large_en_v1_5,       "tokiers/bge-large-en-v1.5",      "BAAI/bge-large-en-v1.5");
accuracy_test!(bge_small_en_v1_5,       "tokiers/bge-small-en-v1.5",      "BAAI/bge-small-en-v1.5");
accuracy_test!(bge_en_icl,              "tokiers/bge-en-icl",             "BAAI/bge-en-icl");
accuracy_test!(e5_base_v2,              "tokiers/e5-base-v2",             "intfloat/e5-base-v2");
accuracy_test!(e5_large_v2,             "tokiers/e5-large-v2",            "intfloat/e5-large-v2");
accuracy_test!(e5_small_v2,             "tokiers/e5-small-v2",            "intfloat/e5-small-v2");
accuracy_test!(gte_base,                "tokiers/gte-base",               "thenlper/gte-base");
accuracy_test!(gte_large,               "tokiers/gte-large",              "thenlper/gte-large");
accuracy_test!(gte_small,               "tokiers/gte-small",              "thenlper/gte-small");
accuracy_test!(gte_qwen2_7b_instruct,   "tokiers/gte-Qwen2-7B-instruct", "Alibaba-NLP/gte-Qwen2-7B-instruct");
accuracy_test!(ms_marco_minilm_l_4_v2,  "tokiers/ms-marco-MiniLM-L-4-v2","cross-encoder/ms-marco-MiniLM-L-4-v2");
accuracy_test!(ms_marco_minilm_l_6_v2,  "tokiers/ms-marco-MiniLM-L-6-v2","cross-encoder/ms-marco-MiniLM-L-6-v2");
accuracy_test!(mxbai_embed_large_v1,    "tokiers/mxbai-embed-large-v1",   "mixedbread-ai/mxbai-embed-large-v1");
accuracy_test!(mxbai_embed_2d_large_v1, "tokiers/mxbai-embed-2d-large-v1","mixedbread-ai/mxbai-embed-2d-large-v1");
accuracy_test!(mxbai_embed_xsmall_v1,   "tokiers/mxbai-embed-xsmall-v1",  "mixedbread-ai/mxbai-embed-xsmall-v1");
accuracy_test!(deepset_mxbai_embed_de,  "tokiers/deepset-mxbai-embed-de-large-v1", "mixedbread-ai/deepset-mxbai-embed-de-large-v1");
accuracy_test!(nomic_embed_text_v1,     "tokiers/nomic-embed-text-v1",    "nomic-ai/nomic-embed-text-v1");

// BPE (byte-level)
accuracy_test!(gpt2,                    "tokiers/gpt2",                   "openai-community/gpt2");
accuracy_test!(roberta_base,            "tokiers/roberta-base",           "FacebookAI/roberta-base");
accuracy_test!(phi_2,                   "tokiers/phi-2",                  "microsoft/phi-2");
accuracy_test!(phi_3_mini,              "tokiers/Phi-3-mini-4k-instruct", "microsoft/Phi-3-mini-4k-instruct");
accuracy_test!(modernbert_base,         "tokiers/ModernBERT-base",        "answerdotai/ModernBERT-base");
accuracy_test!(codellama_7b,            "tokiers/CodeLlama-7b-hf",       "tokiers/CodeLlama-7b-hf");
accuracy_test!(llama_3_2_1b,            "tokiers/Llama-3.2-1B",          "tokiers/Llama-3.2-1B");
accuracy_test!(llama_4_scout,           "tokiers/Llama-4-Scout-17B-16E", "tokiers/Llama-4-Scout-17B-16E");
accuracy_test!(mistral_7b,              "tokiers/Mistral-7B-v0.1",       "mistralai/Mistral-7B-v0.1");
accuracy_test!(mistral_nemo,            "tokiers/Mistral-Nemo-Base-2407","mistralai/Mistral-Nemo-Base-2407");
accuracy_test!(mixtral_8x7b,            "tokiers/Mixtral-8x7B-v0.1",    "mistralai/Mixtral-8x7B-v0.1");
accuracy_test!(qwen2_7b,               "tokiers/Qwen2-7B",              "Qwen/Qwen2-7B");
accuracy_test!(qwen3_embed_0_6b,        "tokiers/Qwen3-Embedding-0.6B",  "Qwen/Qwen3-Embedding-0.6B");
accuracy_test!(qwen3_embed_4b,          "tokiers/Qwen3-Embedding-4B",    "Qwen/Qwen3-Embedding-4B");
accuracy_test!(qwen3_embed_8b,          "tokiers/Qwen3-Embedding-8B",    "Qwen/Qwen3-Embedding-8B");

// Jina (BPE)
accuracy_test!(jina_v2_base_en,         "tokiers/jina-embeddings-v2-base-en",   "jinaai/jina-embeddings-v2-base-en");
accuracy_test!(jina_v2_base_code,       "tokiers/jina-embeddings-v2-base-code", "jinaai/jina-embeddings-v2-base-code");
accuracy_test!(jina_v3,                 "tokiers/jina-embeddings-v3",           "jinaai/jina-embeddings-v3");
accuracy_test!(jina_v4,                 "tokiers/jina-embeddings-v4",           "jinaai/jina-embeddings-v4");

// Cohere (BPE)
accuracy_test!(cohere_english_v3,       "tokiers/Cohere-embed-english-v3.0",            "Cohere/Cohere-embed-english-v3.0");
accuracy_test!(cohere_english_light_v3, "tokiers/Cohere-embed-english-light-v3.0",      "Cohere/Cohere-embed-english-light-v3.0");
accuracy_test!(cohere_multi_v3,         "tokiers/Cohere-embed-multilingual-v3.0",       "Cohere/Cohere-embed-multilingual-v3.0");
accuracy_test!(cohere_multi_light_v3,   "tokiers/Cohere-embed-multilingual-light-v3.0", "Cohere/Cohere-embed-multilingual-light-v3.0");

// Voyage (BPE)
accuracy_test!(voyage_3,                "tokiers/voyage-3",                "voyageai/voyage-3");
accuracy_test!(voyage_3_large,          "tokiers/voyage-3-large",          "voyageai/voyage-3-large");
accuracy_test!(voyage_3_lite,           "tokiers/voyage-3-lite",           "voyageai/voyage-3-lite");
accuracy_test!(voyage_3_5,              "tokiers/voyage-3.5",              "voyageai/voyage-3.5");
accuracy_test!(voyage_3_5_lite,         "tokiers/voyage-3.5-lite",         "voyageai/voyage-3.5-lite");
accuracy_test!(voyage_code_2,           "tokiers/voyage-code-2",           "voyageai/voyage-code-2");
accuracy_test!(voyage_code_3,           "tokiers/voyage-code-3",           "voyageai/voyage-code-3");
accuracy_test!(voyage_finance_2,        "tokiers/voyage-finance-2",        "voyageai/voyage-finance-2");
accuracy_test!(voyage_law_2,            "tokiers/voyage-law-2",            "voyageai/voyage-law-2");
accuracy_test!(voyage_multilingual_2,   "tokiers/voyage-multilingual-2",   "voyageai/voyage-multilingual-2");
accuracy_test!(voyage_multimodal_3,     "tokiers/voyage-multimodal-3",     "voyageai/voyage-multimodal-3");

// SentencePiece / Unigram
accuracy_test!(t5_base,                 "tokiers/t5-base",              "google-t5/t5-base");
accuracy_test!(xlm_roberta_base,        "tokiers/xlm-roberta-base",     "FacebookAI/xlm-roberta-base");

// New models
accuracy_test!(deepseek_v3,             "tokiers/DeepSeek-V3",                      "deepseek-ai/DeepSeek-V3");
accuracy_test!(deepseek_r1,             "tokiers/DeepSeek-R1",                      "deepseek-ai/DeepSeek-R1");
accuracy_test!(gemma_2_2b,              "tokiers/gemma-2-2b",                       "tokiers/gemma-2-2b");
accuracy_test!(gemma_3_4b_it,           "tokiers/gemma-3-4b-it",                    "tokiers/gemma-3-4b-it");
accuracy_test!(bge_m3,                  "tokiers/bge-m3",                           "BAAI/bge-m3");
accuracy_test!(snowflake_arctic_v2,     "tokiers/snowflake-arctic-embed-l-v2.0",    "Snowflake/snowflake-arctic-embed-l-v2.0");
accuracy_test!(nv_embed_v2,             "tokiers/NV-Embed-v2",                      "nvidia/NV-Embed-v2");
accuracy_test!(smollm2_135m,            "tokiers/SmolLM2-135M",                     "HuggingFaceTB/SmolLM2-135M");
accuracy_test!(stablelm_2_1_6b,         "tokiers/stablelm-2-1_6b",                  "stabilityai/stablelm-2-1_6b");

// Qwen3 / Qwen3.5
accuracy_test!(qwen3_0_6b,              "tokiers/Qwen3-0.6B",                       "Qwen/Qwen3-0.6B");
accuracy_test!(qwen3_8b,                "tokiers/Qwen3-8B",                         "Qwen/Qwen3-8B");
accuracy_test!(qwen3_coder_30b,         "tokiers/Qwen3-Coder-30B-A3B-Instruct",     "Qwen/Qwen3-Coder-30B-A3B-Instruct");
accuracy_test!(qwen3_5_0_8b,            "tokiers/Qwen3.5-0.8B",                     "Qwen/Qwen3.5-0.8B");
accuracy_test!(qwen3_5_4b,              "tokiers/Qwen3.5-4B",                       "Qwen/Qwen3.5-4B");

// ============================================================================
// tiktoken models (CL100K, O200K) — compared against tiktoken-rs
// ============================================================================

/// Compare tokie against tiktoken-rs.
fn compare_tiktoken(tokiers_repo: &str, tiktoken_model: &str, text: &str) -> (bool, Option<usize>) {
    let tok = Tokenizer::from_pretrained(tokiers_repo)
        .unwrap_or_else(|e| panic!("Failed to load tokie {tokiers_repo}: {e}"));
    let bpe = tiktoken_rs::get_bpe_from_model(tiktoken_model)
        .unwrap_or_else(|e| panic!("Failed to load tiktoken {tiktoken_model}: {e}"));

    let tokie_ids = tok.encode(text, false).ids;
    let tiktoken_ids = bpe.encode_with_special_tokens(text);

    if tokie_ids.as_slice() == tiktoken_ids.as_slice() {
        (true, None)
    } else {
        let diff = tokie_ids.iter().zip(tiktoken_ids.iter())
            .position(|(a, b)| a != b)
            .unwrap_or(tokie_ids.len().min(tiktoken_ids.len()));
        (false, Some(diff))
    }
}

macro_rules! tiktoken_accuracy_test {
    ($name:ident, $tokiers:expr, $tiktoken_model:expr) => {
        #[test]
        #[ignore]
        fn $name() {
            let text = load_enwik8(1_000_000);
            let (pass, diff) = compare_tiktoken($tokiers, $tiktoken_model, &text);
            assert!(pass, "Token mismatch at index {:?}", diff);
        }
    };
}

tiktoken_accuracy_test!(tiktoken_cl100k, "tokiers/cl100k", "gpt-4");
tiktoken_accuracy_test!(tiktoken_o200k,  "tokiers/o200k",  "gpt-4o");