package saga

  1. Overview
  2. Docs

Source file saga_tokenizers.ml

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
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
(** Main tokenizers module that aggregates all tokenization functionality *)

module Either = struct
  type ('a, 'b) t = Left of 'a | Right of 'b
end

module Models = Models
(** Re-export all the submodules *)

module Normalizers = Normalizers
module Pre_tokenizers = Pre_tokenizers
module Processors = Processors
module Decoders = Decoders
module Trainers = Trainers
module Unicode = Unicode
module Encoding = Encoding
module Bpe = Bpe
module Wordpiece = Wordpiece

(* Compatibility layer implementation *)

type split_delimiter_behavior =
  [ `Removed
  | `Isolated
  | `Merged_with_previous
  | `Merged_with_next
  | `Contiguous ]
(** Type definitions *)

type strategy = [ `Longest_first | `Only_first | `Only_second ]
type prepend_scheme = [ `Always | `Never | `First ]

(** Added token type *)
module Added_token = struct
  type t = {
    content : string;
    single_word : bool;
    lstrip : bool;
    rstrip : bool;
    normalized : bool;
    special : bool;
  }

  let create ?(content = "") ?(single_word = false) ?(lstrip = false)
      ?(rstrip = false) ?(normalized = true) ?(special = false) () =
    { content; single_word; lstrip; rstrip; normalized; special }

  let content t = t.content
  let lstrip t = t.lstrip
  let normalized t = t.normalized
  let rstrip t = t.rstrip
  let single_word t = t.single_word
  let special t = t.special
  let _ = special (* Suppress unused warning *)
end

type direction = [ `Left | `Right ]
(** Direction type for padding *)

(** Main Tokenizer module *)
module Tokenizer = struct
  type padding_config = {
    direction : direction;
    pad_id : int;
    pad_type_id : int;
    pad_token : string;
    length : int option;
    pad_to_multiple_of : int option;
  }

  type truncation_config = {
    max_length : int;
    stride : int;
    strategy : strategy;
    direction : direction;
  }

  type t = {
    mutable model : Models.t;
    mutable normalizer : Normalizers.t option;
    mutable pre_tokenizer : Pre_tokenizers.t option;
    mutable post_processor : Processors.t option;
    mutable decoder : Decoders.t option;
    mutable padding : padding_config option;
    mutable truncation : truncation_config option;
    added_tokens : (string, Added_token.t) Hashtbl.t;
    special_tokens : (string, int) Hashtbl.t;
  }

  let create ~model =
    {
      model;
      normalizer = None;
      pre_tokenizer = None;
      post_processor = None;
      decoder = None;
      padding = None;
      truncation = None;
      added_tokens = Hashtbl.create 16;
      special_tokens = Hashtbl.create 16;
    }

  let from_file path =
    try
      let json = Yojson.Basic.from_file path in
      (* Parse tokenizer.json structure *)
      let t =
        match json with
        | `Assoc fields ->
            let model =
              match List.assoc_opt "model" fields with
              | Some model_json -> Models.of_json model_json
              | None -> Models.word_level ()
            in
            let tok = create ~model in
            (* Optional components *)
            (match List.assoc_opt "normalizer" fields with
            | Some `Null | None -> ()
            | Some njson -> tok.normalizer <- Some (Normalizers.of_json njson));
            (match List.assoc_opt "post_processor" fields with
            | Some `Null | None -> ()
            | Some pjson ->
                tok.post_processor <- Some (Processors.of_json pjson));
            (match List.assoc_opt "truncation" fields with
            | Some (`Assoc tfs) ->
                let max_length =
                  match List.assoc_opt "max_length" tfs with
                  | Some (`Int i) -> i
                  | _ -> 0
                in
                let stride =
                  match List.assoc_opt "stride" tfs with
                  | Some (`Int i) -> i
                  | _ -> 0
                in
                let direction =
                  match List.assoc_opt "direction" tfs with
                  | Some (`String "Left") -> `Left
                  | _ -> `Right
                in
                let strategy =
                  match List.assoc_opt "strategy" tfs with
                  | Some (`String "OnlyFirst") -> `Only_first
                  | Some (`String "OnlySecond") -> `Only_second
                  | _ -> `Longest_first
                in
                tok.truncation <-
                  Some { max_length; stride; strategy; direction }
            | _ -> ());
            (match List.assoc_opt "padding" fields with
            | Some (`Assoc pfs) ->
                let direction =
                  match List.assoc_opt "direction" pfs with
                  | Some (`String "Left") -> `Left
                  | _ -> `Right
                in
                let pad_id =
                  match List.assoc_opt "pad_id" pfs with
                  | Some (`Int i) -> i
                  | _ -> 0
                in
                let pad_type_id =
                  match List.assoc_opt "pad_type_id" pfs with
                  | Some (`Int i) -> i
                  | _ -> 0
                in
                let pad_token =
                  match List.assoc_opt "pad_token" pfs with
                  | Some (`String s) -> s
                  | _ -> "<pad>"
                in
                let length =
                  match List.assoc_opt "length" pfs with
                  | Some (`Int i) -> Some i
                  | _ -> None
                in
                let pad_to_multiple_of =
                  match List.assoc_opt "pad_to_multiple_of" pfs with
                  | Some (`Int i) -> Some i
                  | _ -> None
                in
                tok.padding <-
                  Some
                    {
                      direction;
                      pad_id;
                      pad_type_id;
                      pad_token;
                      length;
                      pad_to_multiple_of;
                    }
            | _ -> ());
            tok
        | _ -> create ~model:(Models.word_level ())
      in
      Ok t
    with e -> Error e

  let from_str str =
    try
      let bytes = Bytes.of_string str in
      let tmp = Filename.temp_file "saga_tokenizer" ".json" in
      let oc = open_out tmp in
      output_string oc (Bytes.to_string bytes);
      close_out oc;
      let res = from_file tmp in
      (try Sys.remove tmp with _ -> ());
      res
    with e -> Error e

  let from_pretrained identifier ?revision ?token () =
    (* Would need to download from HF hub *)
    let _ = (identifier, revision, token) in
    let model = Models.word_level () in
    Ok (create ~model)

  let from_buffer bytes = from_str (Bytes.to_string bytes)

  let save t ~path ?pretty () =
    let _ = pretty in
    (* Create a basic JSON structure *)
    let json =
      `Assoc
        [
          ("version", `String "1.0");
          ("truncation", `Null);
          ("padding", `Null);
          ("added_tokens", `List []);
          ("normalizer", `Null);
          ("pre_tokenizer", `Null);
          ("post_processor", `Null);
          ("decoder", `Null);
          ("model", Models.to_json t.model);
        ]
    in
    let oc = open_out path in
    Yojson.Basic.to_channel oc json;
    close_out oc

  let to_str t ?pretty () =
    let int_opt = function None -> `Null | Some i -> `Int i in
    let json =
      `Assoc
        [
          ("version", `String "1.0");
          ( "truncation",
            match t.truncation with
            | None -> `Null
            | Some cfg ->
                `Assoc
                  [
                    ("max_length", `Int cfg.max_length);
                    ("stride", `Int cfg.stride);
                    ( "strategy",
                      `String
                        (match cfg.strategy with
                        | `Longest_first -> "LongestFirst"
                        | `Only_first -> "OnlyFirst"
                        | `Only_second -> "OnlySecond") );
                    ( "direction",
                      `String
                        (match cfg.direction with
                        | `Left -> "Left"
                        | `Right -> "Right") );
                  ] );
          ( "padding",
            match t.padding with
            | None -> `Null
            | Some cfg ->
                `Assoc
                  [
                    ( "direction",
                      `String
                        (match cfg.direction with
                        | `Left -> "Left"
                        | `Right -> "Right") );
                    ("pad_id", `Int cfg.pad_id);
                    ("pad_type_id", `Int cfg.pad_type_id);
                    ("pad_token", `String cfg.pad_token);
                    ("length", int_opt cfg.length);
                    ("pad_to_multiple_of", int_opt cfg.pad_to_multiple_of);
                  ] );
          ( "added_tokens",
            `List
              (Hashtbl.fold
                 (fun _ tok acc ->
                   `Assoc
                     [
                       ("content", `String (Added_token.content tok));
                       ("single_word", `Bool (Added_token.single_word tok));
                       ("lstrip", `Bool (Added_token.lstrip tok));
                       ("rstrip", `Bool (Added_token.rstrip tok));
                       ("normalized", `Bool (Added_token.normalized tok));
                       ("special", `Bool (Added_token.special tok));
                     ]
                   :: acc)
                 t.added_tokens []) );
          ( "normalizer",
            match t.normalizer with
            | None -> `Null
            | Some n -> Normalizers.to_json n );
          ( "pre_tokenizer",
            match t.pre_tokenizer with
            | None -> `Null
            | Some _ ->
                (* We do not serialize function values; represent symbolic
                   name *)
                `Assoc [ ("type", `String "Custom") ] );
          ( "post_processor",
            match t.post_processor with
            | None -> `Null
            | Some p -> Processors.to_json p );
          ( "decoder",
            match t.decoder with
            | None -> `Null
            | Some _ -> `Assoc [ ("type", `String "Custom") ] );
          ("model", Models.to_json t.model);
        ]
    in
    match pretty with
    | Some true -> Yojson.Basic.pretty_to_string json
    | _ -> Yojson.Basic.to_string json

  let get_model t = t.model
  let set_model t model = t.model <- model
  let get_normalizer t = t.normalizer
  let set_normalizer t normalizer = t.normalizer <- normalizer
  let get_pre_tokenizer t = t.pre_tokenizer
  let set_pre_tokenizer t pre_tokenizer = t.pre_tokenizer <- pre_tokenizer
  let get_post_processor t = t.post_processor
  let set_post_processor t post_processor = t.post_processor <- post_processor
  let get_decoder t = t.decoder
  let set_decoder t decoder = t.decoder <- decoder
  let get_padding t = t.padding
  let _set_padding t padding = t.padding <- Some padding
  let no_padding t = t.padding <- None
  let get_truncation t = t.truncation
  let _set_truncation t truncation = t.truncation <- Some truncation
  let no_truncation t = t.truncation <- None

  let add_special_tokens t tokens =
    let count = ref 0 in
    let token_strings = ref [] in
    List.iter
      (fun token_either ->
        let added_token =
          match token_either with
          | Either.Left str -> Added_token.create ~content:str ~special:true ()
          | Either.Right tok -> tok
        in
        if not (Hashtbl.mem t.added_tokens (Added_token.content added_token))
        then (
          Hashtbl.add t.added_tokens
            (Added_token.content added_token)
            added_token;
          (* TODO: properly assign IDs to special tokens *)
          Hashtbl.add t.special_tokens
            (Added_token.content added_token)
            (Hashtbl.length t.special_tokens);
          token_strings := Added_token.content added_token :: !token_strings;
          incr count))
      tokens;
    (* Also add special tokens to the model's vocabulary *)
    let model_count = Models.add_tokens t.model (List.rev !token_strings) in
    let _ = model_count in
    !count

  let add_tokens t tokens =
    let count = ref 0 in
    let token_strings = ref [] in
    List.iter
      (fun token_either ->
        let added_token =
          match token_either with
          | Either.Left str -> Added_token.create ~content:str ()
          | Either.Right tok -> tok
        in
        if not (Hashtbl.mem t.added_tokens (Added_token.content added_token))
        then (
          Hashtbl.add t.added_tokens
            (Added_token.content added_token)
            added_token;
          token_strings := Added_token.content added_token :: !token_strings;
          incr count))
      tokens;
    (* Also add tokens to the model's vocabulary *)
    let model_count = Models.add_tokens t.model (List.rev !token_strings) in
    let _ = model_count in
    (* Could verify this matches count *)
    !count

  let get_vocab t ?with_added_tokens () =
    let _with_added = Option.value with_added_tokens ~default:true in
    (* Added tokens are injected into model vocab via Models.add_tokens, so
       model vocab is authoritative *)
    Models.get_vocab t.model

  let get_vocab_size t ?with_added_tokens () =
    let with_added = Option.value with_added_tokens ~default:true in
    let base_size = Models.get_vocab_size t.model in
    if with_added then
      base_size
      + Hashtbl.length t.added_tokens
      + Hashtbl.length t.special_tokens
    else base_size

  let enable_padding t config = t.padding <- Some config
  let enable_truncation t config = t.truncation <- Some config

  (* Conversion helpers between Encoding.t and Processors.encoding *)
  let encoding_to_processor (e : Encoding.t) : Processors.encoding =
    {
      ids = Encoding.get_ids e;
      type_ids = Encoding.get_type_ids e;
      tokens = Encoding.get_tokens e;
      offsets = Encoding.get_offsets e;
      special_tokens_mask = Encoding.get_special_tokens_mask e;
      attention_mask = Encoding.get_attention_mask e;
      overflowing = [];
      (* Encoding has overflowing; ignore in processor for now *)
      sequence_ranges = [];
    }

  let processor_to_encoding (pe : Processors.encoding) : Encoding.t =
    (* Rebuild Encoding.t; words are None; sequence ranges empty *)
    let len = Array.length pe.ids in
    let words = Array.make len None in
    let seq = Hashtbl.create 1 in
    Encoding.create ~ids:pe.ids ~type_ids:pe.type_ids ~tokens:pe.tokens ~words
      ~offsets:pe.offsets ~special_tokens_mask:pe.special_tokens_mask
      ~attention_mask:pe.attention_mask ~overflowing:[] ~sequence_ranges:seq

  let encode t ~sequence ?pair ?is_pretokenized ?add_special_tokens () =
    let is_pretokenized = Option.value is_pretokenized ~default:false in
    let add_special_tokens = Option.value add_special_tokens ~default:true in
    let _ = (pair, is_pretokenized, add_special_tokens) in

    (* Handle Either type for sequence *)
    let text =
      match sequence with
      | Either.Left s -> s
      | Either.Right lst -> String.concat " " lst
    in

    (* Normalize *)
    let normalized =
      match t.normalizer with
      | Some n -> Normalizers.normalize_str n text
      | None -> text
    in

    (* Pre-tokenize *)
    let pre_tokens =
      match t.pre_tokenizer with
      | Some pt -> pt normalized (* Pre_tokenizers.t is already a function *)
      | None -> [ (normalized, (0, String.length normalized)) ]
    in

    (* Tokenize *)
    let tokens =
      List.concat_map (fun (text, _) -> Models.tokenize t.model text) pre_tokens
    in

    (* Convert to encoding using the new Encoding module *)
    let token_list =
      List.map
        (fun (tok : Models.token) -> (tok.id, tok.value, tok.offsets))
        tokens
    in
    let enc = Encoding.from_tokens token_list ~type_id:0 in
    (* Post-process if configured *)
    let enc =
      match t.post_processor with
      | None -> enc
      | Some proc -> (
          let proc_enc = encoding_to_processor enc in
          let outs = Processors.process proc [ proc_enc ] ~add_special_tokens in
          match outs with [] -> enc | x :: _ -> processor_to_encoding x)
    in
    (* Apply truncation if configured *)
    let enc =
      match t.truncation with
      | None -> enc
      | Some cfg ->
          let t_dir : Encoding.truncation_direction =
            match cfg.direction with
            | `Left -> Encoding.Left
            | `Right -> Encoding.Right
          in
          Encoding.truncate enc ~max_length:cfg.max_length ~stride:cfg.stride
            ~direction:t_dir
    in
    (* Apply padding if configured and length provided *)
    let enc =
      match t.padding with
      | None -> enc
      | Some cfg -> (
          let p_dir : Encoding.padding_direction =
            match cfg.direction with
            | `Left -> Encoding.Left
            | `Right -> Encoding.Right
          in
          match (cfg.length, cfg.pad_to_multiple_of) with
          | Some target, _ ->
              Encoding.pad enc ~target_length:target ~pad_id:cfg.pad_id
                ~pad_type_id:cfg.pad_type_id ~pad_token:cfg.pad_token
                ~direction:p_dir
          | None, Some m when m > 0 ->
              let cur = Encoding.length enc in
              let target = (cur + m - 1) / m * m in
              Encoding.pad enc ~target_length:target ~pad_id:cfg.pad_id
                ~pad_type_id:cfg.pad_type_id ~pad_token:cfg.pad_token
                ~direction:p_dir
          | None, _ -> enc)
    in
    enc

  let encode_batch t ~input ?is_pretokenized ?add_special_tokens () =
    let encs =
      List.map
        (fun item ->
          match item with
          | Either.Left sequence ->
              encode t ~sequence ?pair:None ?is_pretokenized ?add_special_tokens
                ()
          | Either.Right (seq1, seq2) ->
              encode t ~sequence:seq1 ~pair:seq2 ?is_pretokenized
                ?add_special_tokens ())
        input
    in
    (* If padding enabled without fixed length, pad to the max in the batch *)
    match t.padding with
    | None -> encs
    | Some cfg ->
        let p_dir : Encoding.padding_direction =
          match cfg.direction with
          | `Left -> Encoding.Left
          | `Right -> Encoding.Right
        in
        let max_len =
          List.fold_left (fun acc e -> max acc (Encoding.length e)) 0 encs
        in
        let target =
          match (cfg.length, cfg.pad_to_multiple_of) with
          | Some l, _ -> l
          | None, Some m when m > 0 -> (max_len + m - 1) / m * m
          | _ -> max_len
        in
        List.map
          (fun e ->
            if Encoding.length e >= target then e
            else
              Encoding.pad e ~target_length:target ~pad_id:cfg.pad_id
                ~pad_type_id:cfg.pad_type_id ~pad_token:cfg.pad_token
                ~direction:p_dir)
          encs

  let decode t ids ?skip_special_tokens ?clean_up_tokenization_spaces () =
    let skip_special_tokens = Option.value skip_special_tokens ~default:false in
    let clean_up_tokenization_spaces =
      Option.value clean_up_tokenization_spaces ~default:true
    in
    let _ = clean_up_tokenization_spaces in

    (* Get tokens from IDs and filter special tokens if requested *)
    let tokens =
      List.filter_map
        (fun id ->
          match Models.id_to_token t.model id with
          | Some token
            when skip_special_tokens && Hashtbl.mem t.special_tokens token ->
              None
          | Some token -> Some token
          | None -> None)
        ids
    in

    (* Apply decoder if present *)
    match t.decoder with
    | Some d -> Decoders.decode d tokens
    | None -> (
        (* Default decoding behavior depends on model type *)
        match t.model with
        | Models.WordLevel _ ->
            (* WordLevel adds spaces between tokens by default *)
            String.concat " " tokens
        | Models.WordPiece _ ->
            (* WordPiece concatenates without spaces but handles ## prefix *)
            String.concat "" tokens
        | _ ->
            (* Other models concatenate without spaces *)
            String.concat "" tokens)

  let decode_batch t id_lists ?skip_special_tokens ?clean_up_tokenization_spaces
      () =
    List.map
      (fun ids ->
        decode t ids ?skip_special_tokens ?clean_up_tokenization_spaces ())
      id_lists

  let token_to_id t token = Models.token_to_id t.model token
  let id_to_token t id = Models.id_to_token t.model id

  let get_added_tokens_decoder t =
    (* Return mapping of added token ids to their Added_token.t when present in
       model vocab *)
    Hashtbl.fold
      (fun _ tok acc ->
        match Models.token_to_id t.model (Added_token.content tok) with
        | Some id -> (id, tok) :: acc
        | None -> acc)
      t.added_tokens []

  let train t ~files ?trainer () =
    let trainer =
      match trainer with Some tr -> tr | None -> Trainers.word_level ()
    in
    let res = Trainers.train trainer ~files ?model:(Some t.model) () in
    (* Update model and register special tokens *)
    t.model <- res.model;
    let _ =
      if res.special_tokens <> [] then
        add_special_tokens t
          (List.map (fun s -> Either.Left s) res.special_tokens)
      else 0
    in
    ()

  let train_from_iterator t (seq : string Seq.t) ?trainer ?length () =
    let _ = length in
    let trainer =
      match trainer with Some tr -> tr | None -> Trainers.word_level ()
    in
    (* Convert Seq.t to iterator (unit -> string option) *)
    let state = ref seq in
    let iterator () =
      match !state () with
      | Seq.Nil -> None
      | Seq.Cons (x, next) ->
          state := next;
          Some x
    in
    let res =
      Trainers.train_from_iterator trainer ~iterator ?model:(Some t.model) ()
    in
    t.model <- res.model;
    let _ =
      if res.special_tokens <> [] then
        add_special_tokens t
          (List.map (fun s -> Either.Left s) res.special_tokens)
      else 0
    in
    ()

  let post_process _t ~encoding ?pair ?add_special_tokens () =
    let _add_special_tokens = Option.value add_special_tokens ~default:true in
    let _ = pair in
    (* Not fully wired to Processors.encoding; return identity for now *)
    encoding

  let num_special_tokens_to_add t ~is_pair =
    match t.post_processor with
    | None -> 0
    | Some p -> Processors.added_tokens p ~is_pair

  let save_pretrained t ~path =
    (* Save tokenizer.json and model files *)
    let tok_file = Filename.concat path "tokenizer.json" in
    let oc = open_out tok_file in
    output_string oc (to_str t ~pretty:true ());
    close_out oc;
    let _files = Models.save t.model ~folder:path () in
    ()
end