fix: seq2seq decode returns primary choice, strips slash-separated alternates (#805)

Author: ronaldtse

lib/interscript/ml/byt5_onnx.rb

@@ -24,7 +24,7 @@ def translate(text, max_seq_length: 256)
 
       hidden = @encoder.predict({ input_ids: [ids] })['last_hidden_state']
       tokens = @kv ? greedy_kv(hidden, max_seq_length) : greedy_plain(hidden, max_seq_length)
-      IMF.decode(tokens)
+      IMF.first_alternates(IMF.decode(tokens))
     end
 
     def id

lib/interscript/ml/imf.rb

@@ -28,6 +28,13 @@ def encode(text)
         text.bytes.map { |b| b + BYTE_OFFSET } + [EOS_ID]
       end
 
+      # Seq2seq output contract: models trained on multi-reference
+      # corpora emit '/'-separated alternates per token; the runtime
+      # returns the primary choice only.
+      def first_alternates(text)
+        text.split.map { |tok| tok.split('/', 2).first }.join(' ')
+      end
+
       def decode(token_ids)
         out = +""
         token_ids.each do |token|

spec/first_alternates_spec.rb

@@ -0,0 +1,19 @@
+require "spec_helper"
+require "interscript/ml/imf"
+
+RSpec.describe Interscript::ML::IMF do
+  describe ".first_alternates" do
+    it "keeps the primary choice per token" do
+      expect(described_class.first_alternates("وِكَالَةُ/وِكَالَةُ اَلْأَرْصَادِ")).to eq("وِكَالَةُ اَلْأَرْصَادِ")
+    end
+
+    it "leaves plain text untouched" do
+      expect(described_class.first_alternates("مرحبا بالعالم")).to eq("مرحبا بالعالم")
+    end
+
+    it "handles edges" do
+      expect(described_class.first_alternates("")).to eq("")
+      expect(described_class.first_alternates("a/b c/d")).to eq("a c")
+    end
+  end
+end