diff --git a/compiler_opt/rl/log_reader.py b/compiler_opt/rl/log_reader.py
index e2f63aed..88f0101a 100644
--- a/compiler_opt/rl/log_reader.py
+++ b/compiler_opt/rl/log_reader.py
@@ -61,7 +61,8 @@
 import json
 import math
 
-from typing import Any, BinaryIO, Dict, Generator, List, Optional
+from typing import Any, BinaryIO, Dict, Generator, List, Optional, Union
+import numpy as np
 import tensorflow as tf
 
 _element_type_name_map = {
@@ -86,6 +87,11 @@
 }
 
 
+def convert_dtype_to_ctype(dtype: str) -> Union[type, tf.dtypes.DType]:
+  """Public interface for the _dtype_to_ctype dict."""
+  return _dtype_to_ctype[dtype]
+
+
 def create_tensorspec(d: Dict[str, Any]) -> tf.TensorSpec:
   name: str = d['name']
   shape: List[int] = [int(e) for e in d['shape']]
@@ -120,6 +126,12 @@ def __init__(self, spec: tf.TensorSpec, buffer: bytes):
   def spec(self):
     return self._spec
 
+  def to_numpy(self) -> np.ndarray:
+    return np.frombuffer(
+        self._buffer,
+        dtype=convert_dtype_to_ctype(self._spec.dtype),
+        count=self._len)
+
   def _set_view(self):
     # c_char_p is a nul-terminated string, so the more appropriate cast here
     # would be POINTER(c_char), but unfortunately, c_char_p is the only
@@ -205,11 +217,15 @@ def _enumerate_log_from_stream(
         score=score)
 
 
+def read_log_from_file(f) -> Generator[ObservationRecord, None, None]:
+  header = _read_header(f)
+  if header:
+    yield from _enumerate_log_from_stream(f, header)
+
+
 def read_log(fname: str) -> Generator[ObservationRecord, None, None]:
   with open(fname, 'rb') as f:
-    header = _read_header(f)
-    if header:
-      yield from _enumerate_log_from_stream(f, header)
+    yield from read_log_from_file(f)
 
 
 def _add_feature(se: tf.train.SequenceExample, spec: tf.TensorSpec,
diff --git a/compiler_opt/rl/log_reader_test.py b/compiler_opt/rl/log_reader_test.py
index e093ff35..467eeec0 100644
--- a/compiler_opt/rl/log_reader_test.py
+++ b/compiler_opt/rl/log_reader_test.py
@@ -22,6 +22,7 @@
 from google.protobuf import text_format  # pytype: disable=pyi-error
 from typing import BinaryIO
 
+import numpy as np
 import tensorflow as tf
 
 
@@ -38,26 +39,25 @@ def write_buff(f: BinaryIO, buffer: list, ct):
 
 
 def write_context_marker(f: BinaryIO, name: str):
-  f.write(nl)
   f.write(json_to_bytes({'context': name}))
+  f.write(nl)
 
 
 def write_observation_marker(f: BinaryIO, obs_idx: int):
-  f.write(nl)
   f.write(json_to_bytes({'observation': obs_idx}))
+  f.write(nl)
 
 
-def begin_features(f: BinaryIO):
+def write_nl(f: BinaryIO):
   f.write(nl)
 
 
 def write_outcome_marker(f: BinaryIO, obs_idx: int):
-  f.write(nl)
   f.write(json_to_bytes({'outcome': obs_idx}))
+  f.write(nl)
 
 
 def create_example(fname: str, nr_contexts=1):
-
   t0_val = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6]
   t1_val = [1, 2, 3]
   s = [1.2]
@@ -69,12 +69,12 @@ def create_example(fname: str, nr_contexts=1):
                 'name': 'tensor_name2',
                 'port': 0,
                 'shape': [2, 3],
-                'type': 'float'
+                'type': 'float',
             }, {
                 'name': 'tensor_name1',
                 'port': 0,
                 'shape': [3, 1],
-                'type': 'int64_t'
+                'type': 'int64_t',
             }],
             'score': {
                 'name': 'reward',
@@ -83,29 +83,30 @@ def create_example(fname: str, nr_contexts=1):
                 'type': 'float'
             }
         }))
+    write_nl(f)
     for ctx_id in range(nr_contexts):
       t0_val = [v + ctx_id * 10 for v in t0_val]
       t1_val = [v + ctx_id * 10 for v in t1_val]
       write_context_marker(f, f'context_nr_{ctx_id}')
       write_observation_marker(f, 0)
-      begin_features(f)
       write_buff(f, t0_val, ctypes.c_float)
       write_buff(f, t1_val, ctypes.c_int64)
+      write_nl(f)
       write_outcome_marker(f, 0)
-      begin_features(f)
       write_buff(f, s, ctypes.c_float)
+      write_nl(f)
 
       t0_val = [v + 1 for v in t0_val]
       t1_val = [v + 1 for v in t1_val]
       s[0] += 1
 
       write_observation_marker(f, 1)
-      begin_features(f)
       write_buff(f, t0_val, ctypes.c_float)
       write_buff(f, t1_val, ctypes.c_int64)
+      write_nl(f)
       write_outcome_marker(f, 1)
-      begin_features(f)
       write_buff(f, s, ctypes.c_float)
+      write_nl(f)
 
 
 class LogReaderTest(tf.test.TestCase):
@@ -155,6 +156,19 @@ def test_read_log(self):
       obs_id += 1
     self.assertEqual(obs_id, 2)
 
+  def test_to_numpy(self):
+    logfile = self.create_tempfile()
+    create_example(logfile)
+    t0_val = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6]
+    t1_val = [1, 2, 3]
+    for record in log_reader.read_log(logfile):
+      np.testing.assert_allclose(record.feature_values[0].to_numpy(),
+                                 np.array(t0_val))
+      np.testing.assert_allclose(record.feature_values[1].to_numpy(),
+                                 np.array(t1_val))
+      t0_val = [v + 1 for v in t0_val]
+      t1_val = [v + 1 for v in t1_val]
+
   def test_seq_example_conversion(self):
     logfile = self.create_tempfile()
     create_example(logfile, nr_contexts=2)