diff --git a/lingvo/core/ops/BUILD b/lingvo/core/ops/BUILD
index afa48f2ce..73c2571bc 100644
--- a/lingvo/core/ops/BUILD
+++ b/lingvo/core/ops/BUILD
@@ -107,6 +107,14 @@ lingvo_cc_library(
     hdrs = ["simple_vocab.h"],
 )
 
+lingvo_cc_test(
+    name = "simple_vocab_cc_test",
+    srcs = ["simple_vocab_cc_test.cc"],
+    deps = [
+        ":simple_vocab",
+    ],
+)
+
 custom_kernel_library(
     name = "ml_perf_subword_op",
     srcs = ["ml_perf_subword_op.cc"],
diff --git a/lingvo/core/ops/simple_vocab.cc b/lingvo/core/ops/simple_vocab.cc
index 48b0e2cb3..e54d2f1c5 100644
--- a/lingvo/core/ops/simple_vocab.cc
+++ b/lingvo/core/ops/simple_vocab.cc
@@ -15,6 +15,8 @@ limitations under the License.
 
 #include "lingvo/core/ops/simple_vocab.h"
 
+#include <algorithm>
+
 #include "tensorflow/core/framework/op_kernel.h"
 #include "tensorflow/core/framework/tensor.h"
 #include "tensorflow/core/lib/strings/str_util.h"
@@ -80,6 +82,7 @@ Status Vocab::Load(const string& vocab_glob, bool load_token_ids) {
 Status Vocab::Load(const std::vector<string>& lines, bool load_token_ids) {
   id_to_token_.clear();
   token_to_id_.clear();
+  max_token_id_ = -1;
   int32 next_id = 0;
   for (StringPiece line : lines) {
     if (line.empty()) continue;
@@ -89,12 +92,20 @@ Status Vocab::Load(const std::vector<string>& lines, bool load_token_ids) {
     if (!load_token_ids) {
       token_to_id_[tok] = next_id;
       id_to_token_[next_id] = tok;
+      max_token_id_ = std::max(max_token_id_, next_id);
       next_id++;
     } else {
       CHECK_GE(parts.size(), 2);
       const int32 id = std::stoi(parts[1]);
+      if (id < 0 && str_util::StartsWith(tok, kBowStr)) {
+        return errors::InvalidArgument("BOW token ", tok,
+                                       " has negative id ", id, ".");
+      }
       token_to_id_[tok] = id;
       id_to_token_[id] = tok;
+      if (id >= 0) {
+        max_token_id_ = std::max(max_token_id_, id);
+      }
     }
     VLOG(2) << "Vocab " << token_to_id_[tok] << " " << tok;
   }
diff --git a/lingvo/core/ops/simple_vocab.h b/lingvo/core/ops/simple_vocab.h
index 41bc8f6f3..ae3fa15bc 100644
--- a/lingvo/core/ops/simple_vocab.h
+++ b/lingvo/core/ops/simple_vocab.h
@@ -17,6 +17,8 @@ limitations under the License.
 #define LINGVO_CORE_OPS_SIMPLE_VOCAB_H_
 // TODO(zhifengc): Add comments for this class.
 
+#include <cstddef>
+#include <cstring>
 #include <string>
 #include <unordered_map>
 #include <vector>
@@ -110,10 +112,12 @@ class Vocab {
   }
 
   std::vector<bool> GetBowTokenIds() const {
-    std::vector<bool> is_bow_token_id(id_to_token_.size(), false);
+    const size_t num_token_ids =
+        max_token_id_ < 0 ? 0 : static_cast<size_t>(max_token_id_) + 1;
+    std::vector<bool> is_bow_token_id(num_token_ids, false);
     static const int bowStrLen = strlen(kBowStr);
     for (auto const& kv : id_to_token_) {
-      if (kv.second.substr(0, bowStrLen) == kBowStr) {
+      if (kv.first >= 0 && kv.second.substr(0, bowStrLen) == kBowStr) {
         is_bow_token_id[kv.first] = true;
       }
     }
@@ -126,6 +130,7 @@ class Vocab {
   int32 unk_id_ = -1;
   int32 sow_id_ = -1;
   int32 eow_id_ = -1;
+  int32 max_token_id_ = -1;
   bool use_upper_token_symbols_ = false;
   std::unordered_map<int32, string> id_to_token_;
   std::unordered_map<string, int32> token_to_id_;
diff --git a/lingvo/core/ops/simple_vocab_cc_test.cc b/lingvo/core/ops/simple_vocab_cc_test.cc
new file mode 100644
index 000000000..3319dec00
--- /dev/null
+++ b/lingvo/core/ops/simple_vocab_cc_test.cc
@@ -0,0 +1,54 @@
+/* Copyright 2026 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "lingvo/core/ops/simple_vocab.h"
+
+#include <string>
+#include <vector>
+
+#include <gtest/gtest.h>
+
+namespace tensorflow {
+namespace lingvo {
+namespace {
+
+TEST(SimpleVocabTest, GetBowTokenIdsHandlesSparseIds) {
+  Vocab vocab;
+  ASSERT_TRUE(vocab
+                  .Load({"<S>\t3", "</S>\t5", "<UNK>\t7", "plain\t-1",
+                         std::string(kBowStr) + "hello\t100"},
+                        true)
+                  .ok());
+
+  const std::vector<bool> bow_token_ids = vocab.GetBowTokenIds();
+
+  ASSERT_EQ(101, bow_token_ids.size());
+  EXPECT_TRUE(bow_token_ids[100]);
+  EXPECT_FALSE(bow_token_ids[3]);
+  EXPECT_FALSE(bow_token_ids[7]);
+}
+
+TEST(SimpleVocabTest, LoadRejectsNegativeBowTokenIds) {
+  Vocab vocab;
+  EXPECT_FALSE(vocab
+                   .Load({"<S>\t3", "</S>\t5", "<UNK>\t7",
+                          std::string(kBowStr) + "hello\t-1"},
+                         true)
+                   .ok());
+}
+
+}  // namespace
+}  // namespace lingvo
+}  // namespace tensorflow