Support "fuse unknown" configuration (#117)

* Bring over hf token envvar from preview branch

* Add tests for Gemma, including edge cases

Edge cases also added for other BPE tokenizers, but not for T5 yet.

* Sort added tokens by length (descending) to avoid early partial matches

Similar to huggingface/transformers.js@c305c38

* Store vocab as NSString to allow multiple tokens with the same Unicode
canonical representation.

* Remove comments

* Go back to making vocab dictionaries private

* Use ungated copy of Gemma tokenizer

* Use NSString in UnigramTokenizer

* Support fuse_unk configuration.

Tested on one of the T5 testEdgeCases().

* Remove duplicate function

Author: pcuenca

Sources/Tokenizers/BPETokenizer.swift

@@ -45,6 +45,8 @@ class BPETokenizer: PreTrainedTokenizerModel {
     public let unknownToken: String?
     public let unknownTokenId: Int?
 
+    public let fuseUnknownTokens: Bool
+
     required init(tokenizerConfig: Config, tokenizerData: Config, addedTokens: [String : Int]) throws {
         guard let merges = tokenizerData.model?.merges?.value as? [String] else { fatalError("BPETokenizer requires merges") }
         guard let vocab = tokenizerData.model?.vocab?.dictionary as? [NSString: Int] else {
@@ -75,6 +77,8 @@ class BPETokenizer: PreTrainedTokenizerModel {
 
         bosToken = tokenizerConfig.bosToken?.stringValue
         bosTokenId = bosToken == nil ? nil : tokensToIds[bosToken! as NSString]
+
+        fuseUnknownTokens = tokenizerConfig.fuseUnk?.boolValue ?? false
     }
 
     func convertTokenToId(_ token: String) -> Int? {

Sources/Tokenizers/BertTokenizer.swift

@@ -23,11 +23,14 @@ public class BertTokenizer {
     public var eosToken: String?
     public var eosTokenId: Int?
 
+    public let fuseUnknownTokens: Bool
+
     public init(vocab: [String: Int],
-         merges: [String]?,
-         tokenizeChineseChars: Bool = true,
-         bosToken: String? = nil,
-         eosToken: String? = nil
+                merges: [String]?,
+                tokenizeChineseChars: Bool = true,
+                bosToken: String? = nil,
+                eosToken: String? = nil,
+                fuseUnknownTokens: Bool = false
     ) {
         self.vocab = vocab
         self.ids_to_tokens = Utils.invert(vocab)
@@ -37,6 +40,7 @@ public class BertTokenizer {
         self.bosTokenId = bosToken == nil ? nil : vocab[bosToken!]
         self.eosToken = eosToken
         self.eosTokenId = eosToken == nil ? nil : vocab[eosToken!]
+        self.fuseUnknownTokens = fuseUnknownTokens
     }
 
     public required convenience init(tokenizerConfig: Config, tokenizerData: Config, addedTokens: [String : Int]) throws {
@@ -47,7 +51,8 @@ public class BertTokenizer {
         let tokenizeChineseChars = tokenizerConfig.handleChineseChars?.boolValue ?? true
         let eosToken = tokenizerConfig.eosToken?.stringValue
         let bosToken = tokenizerConfig.bosToken?.stringValue
-        self.init(vocab: vocab, merges: merges, tokenizeChineseChars: tokenizeChineseChars, bosToken: bosToken, eosToken: eosToken)
+        let fuseUnknown = tokenizerConfig.fuseUnk?.boolValue ?? false
+        self.init(vocab: vocab, merges: merges, tokenizeChineseChars: tokenizeChineseChars, bosToken: bosToken, eosToken: eosToken, fuseUnknownTokens: fuseUnknown)
     }
 
 

Sources/Tokenizers/Tokenizer.swift

@@ -36,6 +36,8 @@ public protocol TokenizingModel {
     var eosTokenId: Int? { get }
     var unknownToken: String? { get }
     var unknownTokenId: Int? { get }
+
+    var fuseUnknownTokens: Bool { get }
 }
 
 public extension TokenizingModel {
@@ -138,6 +140,7 @@ public class PreTrainedTokenizer: Tokenizer {
     public var eosTokenId: Int? { model.eosTokenId }
     public var unknownToken: String? { model.unknownToken }
     public var unknownTokenId: Int? { model.unknownTokenId }
+    public var fuseUnknownTokens: Bool { model.fuseUnknownTokens }
 
     private let addedTokens: Set<String>
     private let specialTokens: [String: Int]
@@ -232,6 +235,21 @@ public class PreTrainedTokenizer: Tokenizer {
             .replacingOccurrences(of: " 're", with: "'re")
     }
 
+    func fuseUnknown(_ tokens: [String]) -> [String] {
+        guard fuseUnknownTokens else { return tokens }
+        let (fused, _) = tokens.reduce((fused: [String](), previousIsUnknown: false)) { result, token in
+            var (fused, previousIsUnknown) = result
+            let isUnknown = model.convertTokenToId(token) == model.unknownTokenId
+            if isUnknown {
+                if !previousIsUnknown { fused.append(token) }
+            } else {
+                fused.append(token)
+            }
+            return (fused, isUnknown)
+        }
+        return fused
+    }
+
     public func tokenize(text: String) -> [String] {
         // Take care of special tokens first
         let sections: [String]
@@ -243,7 +261,7 @@ public class PreTrainedTokenizer: Tokenizer {
         return sections.enumerated().map { section, x in
             if addedTokens.contains(x) { return [x] }
             return preTokenize(normalize(x), options: section == 0 ? [.firstSection] : []).flatMap { model($0) }
-        }.flatMap { $0 }
+        }.flatMap { fuseUnknown($0) }
     }
 
     /// Main entry point

Sources/Tokenizers/UnigramTokenizer.swift

@@ -29,7 +29,10 @@ class UnigramTokenizer: PreTrainedTokenizerModel {
     let bosTokenId: Int?
     let eosToken: String?
     let eosTokenId: Int?
-    
+
+    // Hardcoded in Unigram tokenizers
+    let fuseUnknownTokens: Bool = true
+
     private let trie: Trie<Character>
 
     required init(tokenizerConfig: Config, tokenizerData: Config, addedTokens: [String : Int]) throws {
@@ -71,10 +74,8 @@ class UnigramTokenizer: PreTrainedTokenizerModel {
 
         trie = Trie()
         trie.append(contentsOf: vocab.map { $0.token })
-                
-        // TODO: set fuse_unk to true
     }
-    
+
     func convertTokenToId(_ token: String) -> Int? {
         return tokensToIds[token as NSString] ?? self.unknownTokenId
     }