diff --git a/instruct_qa/evaluation/metrics.py b/instruct_qa/evaluation/metrics.py
index 65136ba..d829547 100644
--- a/instruct_qa/evaluation/metrics.py
+++ b/instruct_qa/evaluation/metrics.py
@@ -22,9 +22,6 @@
     AutoModelWithLMHead,
     AutoModelForQuestionAnswering,
 )
-import tensorflow as tf
-import tensorflow_hub as hub
-import tensorflow_text as text
 
 import numpy as np
 from scipy.special import softmax
@@ -80,81 +77,6 @@ def __call__(self, predictions, references, questions=None, ids=None):
         }
 
 
-class BEMScore(Metric):
-    def __init__(self, name, **kwargs):
-        super().__init__(name, **kwargs)
-        vocab_path = "gs://cloud-tpu-checkpoints/bert/keras_bert/uncased_L-12_H-768_A-12/vocab.txt"
-        vocab_table = tf.lookup.StaticVocabularyTable(
-            tf.lookup.TextFileInitializer(
-                filename=vocab_path,
-                key_dtype=tf.string,
-                key_index=tf.lookup.TextFileIndex.WHOLE_LINE,
-                value_dtype=tf.int64,
-                value_index=tf.lookup.TextFileIndex.LINE_NUMBER,
-            ),
-            num_oov_buckets=1,
-        )
-        self.cls_id, self.sep_id = vocab_table.lookup(
-            tf.convert_to_tensor(["[CLS]", "[SEP]"])
-        )
-        self.tokenizer = text.BertTokenizer(
-            vocab_lookup_table=vocab_table,
-            token_out_type=tf.int64,
-            preserve_unused_token=True,
-            lower_case=True,
-        )
-        self.bem = hub.load("https://tfhub.dev/google/answer_equivalence/bem/1")
-
-    def bertify_example(self, example):
-        question = self.tokenizer.tokenize(example["question"]).merge_dims(1, 2)
-        reference = self.tokenizer.tokenize(example["reference"]).merge_dims(1, 2)
-        candidate = self.tokenizer.tokenize(example["candidate"]).merge_dims(1, 2)
-        input_ids, segment_ids = text.combine_segments(
-            (candidate, reference, question), self.cls_id, self.sep_id
-        )
-        return {"input_ids": input_ids.numpy(), "segment_ids": segment_ids.numpy()}
-
-    def bertify_examples(self, examples):
-        input_ids = []
-        segment_ids = []
-        for example in examples:
-            example_inputs = self.bertify_example(example)
-            input_ids.append(self.pad(example_inputs["input_ids"]))
-            segment_ids.append(self.pad(example_inputs["segment_ids"]))
-
-        return {"input_ids": np.stack(input_ids), "segment_ids": np.stack(segment_ids)}
-
-    def pad(self, a, length=512):
-        if a.shape[-1] >= length:
-            return a[0][:length]
-        else:
-            return np.append(a, np.zeros(length - a.shape[-1], np.int32))
-
-    def __call__(self, predictions, references, questions, ids=None):
-        assert len(predictions) == len(references)
-
-        scores = []
-        for i in tqdm(range(len(predictions))):
-            examples = [
-                {
-                    "question": questions[i],
-                    "reference": reference,
-                    "candidate": predictions[i],
-                }
-                for reference in references[i]
-            ]
-            inputs = self.bertify_examples(examples)
-            raw_outputs = self.bem(inputs)
-            score = float(max(softmax(raw_outputs, axis=1)[:, 1]))
-            scores.append(score)
-
-        if self.store_individual_scores:
-            individual_scores = [{"bem": score} for score in scores]
-            self.save_individual_scores(ids, individual_scores)
-
-        return {"bem": np.mean(scores)}
-
-
 class Rouge(Metric):
     def __init__(self, name, **kwargs):
         super().__init__(name, **kwargs)