langfun 0.1.2.dev202501010804__py3-none-any.whl → 0.1.2.dev202501030804__py3-none-any.whl

langfun/core/__init__.py

@@ -123,10 +123,6 @@ from langfun.core.memory import Memory
 # Utility for console output.
 from langfun.core import console
 
-# Helpers for implementing _repr_xxx_ methods.
-from langfun.core import repr_utils
-Html = repr_utils.Html
-
 # Utility for event logging.
 from langfun.core import logging
 

langfun/core/eval/matching.py

@@ -251,14 +251,14 @@ class Matching(base.Evaluation):
     for i, (_, example, output, message) in enumerate(self.matches):
       bgcolor = 'white' if i % 2 == 0 else '#DDDDDD'
       s.write(f'<tr style="background-color: {bgcolor}"><td>{i + 1}</td>')
-      input_str = lf.repr_utils.escape_quoted(
+      input_str = pg.Html.escape(
           pg.format(
               example, verbose=False, max_bytes_len=32,
               custom_format=_maybe_html
           )
       )
       s.write(f'<td style="color:green;white-space:pre-wrap">{input_str}</td>')
-      output_str = lf.repr_utils.escape_quoted(
+      output_str = pg.Html.escape(
           pg.format(
               output, verbose=False, max_bytes_len=32,
               custom_format=_maybe_html

langfun/core/eval/scoring.py

@@ -194,9 +194,13 @@ class Scoring(base.Evaluation):
     for i, (example, output, score, message) in enumerate(self.scored):
       bgcolor = 'white' if i % 2 == 0 else '#DDDDDD'
       s.write(f'<tr style="background-color: {bgcolor}"><td>{i + 1}</td>')
-      input_str = pg.format(example, verbose=False, max_bytes_len=32)
+      input_str = pg.Html.escape(
+          pg.format(example, verbose=False, max_bytes_len=32)
+      )
       s.write(f'<td style="color:green;white-space:pre-wrap">{input_str}</td>')
-      output_str = pg.format(output, verbose=False, max_bytes_len=32)
+      output_str = pg.Html.escape(
+          pg.format(output, verbose=False, max_bytes_len=32)
+      )
       s.write(f'<td style="color:blue;white-space:pre-wrap">{output_str}</td>')
       s.write(f'<td style="color:magenta;white-space:pre-wrap">{score}</td>')
       s.write('<td>')

langfun/core/eval/v2/checkpointing.py

@@ -13,8 +13,10 @@
 # limitations under the License.
 """Checkpointing evaluation runs."""
 import abc
+import re
 import threading
 import traceback
+from typing import Annotated
 
 import langfun.core as lf
 from langfun.core.eval.v2 import example as example_lib
@@ -29,6 +31,11 @@ Runner = experiment_lib.Runner
 class Checkpointer(experiment_lib.Plugin):
   """Base class for checkpointing evaluation examples."""
 
+  checkpoint_filename: Annotated[
+      str,
+      'Checkpoint file pattern.'
+  ] = 'checkpoint.bagz'
+
   def on_experiment_start(
       self,
       runner: Runner,
@@ -37,37 +44,35 @@ class Checkpointer(experiment_lib.Plugin):
     if not experiment.is_leaf:
       return
 
-    # For refresh runs, we don't want to load the previous state.
-    if not runner.current_run.refresh:
-      if runner.current_run.input_root != runner.current_run.output_root:
+    current_run = runner.current_run
+    if current_run.reprocess is not True:  # pylint: disable=g-bool-id-comparison
+      if current_run.input_root != current_run.output_root:
         experiment.info(
-            f'Warm starting from directory: {runner.current_run.input_root}.'
+            f'Warm starting from directory: {current_run.input_root}.'
         )
       self._load_experiment(runner, experiment)
 
+    example_ids_to_evaluate = current_run.examples_to_evaluate(experiment)
     if experiment.state.evaluated_examples:
       loaded_example_ids = list(
           sorted(experiment.state.evaluated_examples.keys())
       )
-      example_ids_to_evaluate = (
-          set(runner.current_run.example_ids) if runner.current_run.example_ids
-          else set(range(1, experiment.num_examples + 1))
-      )
       example_ids_to_evaluate -= set(loaded_example_ids)
-
+      example_ids_to_evaluate = list(sorted(example_ids_to_evaluate))
       experiment.info(
-          f'{len(experiment.state.evaluated_examples)} examples have been '
+          f'{len(experiment.state.evaluated_examples)} examples '
           'loaded from checkpoint files. Their outputs will be used '
-          f'for recomputing metrics. Example IDs: {loaded_example_ids}'
+          f'for recomputing metrics. Example IDs: {loaded_example_ids}.'
       )
       experiment.info(
           f'{len(example_ids_to_evaluate)} examples will be processed from '
-          f'scratch. Example IDs: {list(sorted(example_ids_to_evaluate))}'
+          f'scratch. Example IDs: {example_ids_to_evaluate}.'
       )
     else:
       experiment.info(
           'No examples are loaded from checkpoint files. '
-          f'Experiment {experiment.id} starts from scratch.'
+          f'{len(example_ids_to_evaluate)} examples will be processed from '
+          f'scratch. Example IDs: {example_ids_to_evaluate}.'
       )
 
   def on_example_complete(
@@ -81,60 +86,36 @@ class Checkpointer(experiment_lib.Plugin):
       experiment.warning(
           f'Example {example.id} has error. Skipping checkpointing.'
       )
-    else:
+    elif example.newly_processed:
       self._save_example(runner, experiment, example)
 
-  @abc.abstractmethod
-  def _load_experiment(self, runner: Runner, experiment: Experiment) -> None:
-    """Loads the experiment state from checkpoint files."""
-
-  @abc.abstractmethod
-  def _save_example(
-      self,
-      runner: Runner,
-      experiment: Experiment,
-      example: Example,
-  ) -> None:
-    """Saves an evaluated example."""
-
-
-class PerExampleCheckpointer(Checkpointer):
-  """Checkpointer that saves each example to a separate file."""
-
-  checkpoint_filename: str = 'checkpoint.bagz'
-
-  def _on_bound(self):
-    super()._on_bound()
-    prefix, ext = self._file_prefix_and_ext(self.checkpoint_filename)
-    self._checkpoint_file_prefix = prefix
-    self._checkpoint_file_ext = ext
-
   def _load_experiment(
       self,
       runner: Runner,
       experiment: Experiment,
   ) -> None:
     """Creates the checkpoint file."""
-    experiment_dir = runner.current_run.input_dir(experiment)
-    if pg.io.path_exists(experiment_dir):
-      ckpt_files = [
-          runner.current_run.input_path_for(experiment, filename)
-          for filename in pg.io.listdir(experiment_dir)
-          if filename.startswith(self._checkpoint_file_prefix)
-          and filename.endswith(self._checkpoint_file_ext)
-      ]
-    else:
-      ckpt_files = []
-
+    ckpt_files = self._list_checkpoint_filenames(runner, experiment)
     experiment.info(f'Found {len(ckpt_files)} checkpoint files to load.')
 
     # Load the checkpoint files in parallel.
+    current_run = runner.current_run
+    examples_to_load = current_run.examples_to_load(experiment)
+    examples_to_load_metadata = current_run.examples_to_load_metadata(
+        experiment
+    )
     context = dict(counter=0, counter_lock=threading.Lock())
+    copy_ckpt = current_run.input_root != current_run.output_root
+
     def _load_state(ckpt_file):
       error = None
       with pg.timeit() as t:
         try:
-          experiment.load_state(ckpt_file)
+          experiment.load_state(
+              current_run.input_path_for(experiment, ckpt_file),
+              filter=lambda x: x.id in examples_to_load,
+              load_example_metadata=lambda x: x.id in examples_to_load_metadata,
+          )
         except BaseException as e:  # pylint: disable=broad-except
           error = e
         finally:
@@ -144,21 +125,80 @@ class PerExampleCheckpointer(Checkpointer):
           progress_str = f'{context["counter"]}/{len(ckpt_files)}'
           if error is None:
             experiment.info(
-                f'Loaded checkpoint file {ckpt_file} in {t.elapse:.2f} '
+                f'Checkpoint file {ckpt_file!r} loaded in {t.elapse:.2f} '
                 f'seconds. ({progress_str})'
             )
           else:
             experiment.warning(
-                f'Failed to load checkpoint file {ckpt_file}: {error}. '
+                f'Failed to load checkpoint file {ckpt_file!r}: {error}. '
                 f'Skipping the file. ({progress_str})'
             )
 
+        if not copy_ckpt:
+          return
+
+        # Copy the checkpoint records to the output directory.
+        try:
+          with pg.io.open_sequence(
+              current_run.output_path_for(experiment, ckpt_file), 'w'
+          ) as o, pg.io.open_sequence(
+              current_run.input_path_for(experiment, ckpt_file), 'r'
+          ) as i:
+            for x in i:
+              o.add(x)
+        except BaseException as e:  # pylint: disable=broad-except
+          experiment.warning(
+              f'Failed to copy checkpoint {ckpt_file!r}: {e}.'
+          )
+
     _ = list(
         lf.concurrent_map(
             _load_state, ckpt_files, max_workers=16, silence_on_errors=None
         )
     )
 
+  @abc.abstractmethod
+  def _list_checkpoint_filenames(
+      self, runner: Runner, experiment: Experiment
+  ) -> list[str]:
+    """Lists the checkpoint filenames to restore."""
+
+  @abc.abstractmethod
+  def _save_example(
+      self,
+      runner: Runner,
+      experiment: Experiment,
+      example: Example,
+  ) -> None:
+    """Saves an evaluated example."""
+
+
+class PerExampleCheckpointer(Checkpointer):
+  """Checkpointer that saves each example to a separate file."""
+
+  def _on_bound(self):
+    super()._on_bound()
+    prefix, ext = self._file_prefix_and_ext(self.checkpoint_filename)
+    self._checkpoint_file_prefix = prefix
+    self._checkpoint_file_ext = ext
+
+  def _list_checkpoint_filenames(
+      self, runner: Runner, experiment: Experiment
+  ) -> list[str]:
+    experiment_dir = runner.current_run.input_dir(experiment)
+    filenames = []
+    examples_to_load = runner.current_run.examples_to_load(experiment)
+    if pg.io.path_exists(experiment_dir):
+      regex = re.compile(
+          f'{self._checkpoint_file_prefix}_(\\d+){self._checkpoint_file_ext}'
+          .replace('.', '\\.')
+      )
+      for filename in pg.io.listdir(experiment_dir):
+        match = regex.match(filename)
+        if match and int(match.group(1)) in examples_to_load:
+          filenames.append(filename)
+    return filenames
+
   def _save_example(
       self,
       runner: Runner,
@@ -180,11 +220,11 @@ class PerExampleCheckpointer(Checkpointer):
         writer.add(example)
         writer.close()
         experiment.info(
-            f'Example {example.id} saved to {writer.path}.',
+            f'Example {example.id} checkpointed to {writer.path}.',
         )
       except BaseException as e:  # pylint: disable=broad-except
         experiment.error(
-            f'Failed to save example {example.id} to {writer.path}. '
+            f'Failed to checkpoint example {example.id} to {writer.path}. '
             f'Error: {e}, Stacktrace: \n{traceback.format_exc()}.',
         )
         raise e
@@ -201,8 +241,6 @@ class PerExampleCheckpointer(Checkpointer):
 class BulkCheckpointer(Checkpointer):
   """Checkpointer that saves all examples to a single file."""
 
-  checkpoint_filename: str = 'checkpoint.bagz'
-
   def _on_bound(self):
     super()._on_bound()
     self._lock = threading.Lock()
@@ -253,18 +291,14 @@ class BulkCheckpointer(Checkpointer):
         if self._sequence_writer is not None:
           self._sequence_writer[experiment.id] = sequence_writer
 
-  def _load_experiment(
-      self,
-      runner: Runner,
-      experiment: Experiment,
-  ) -> None:
-    """Creates the checkpoint file."""
-    experiment.load_state(
-        runner.current_run.input_path_for(
-            experiment, self.checkpoint_filename
-        ),
-        raise_if_not_exist=False
-    )
+  def _list_checkpoint_filenames(
+      self, runner: Runner, experiment: Experiment
+  ) -> list[str]:
+    if pg.io.path_exists(
+        runner.current_run.input_path_for(experiment, self.checkpoint_filename)
+    ):
+      return [self.checkpoint_filename]
+    return []
 
   def on_experiment_complete(
       self,
@@ -299,11 +333,11 @@ class BulkCheckpointer(Checkpointer):
       try:
         writer.add(example)
         experiment.info(
-            f'Example {example.id} added to {writer.path}.',
+            f'Example {example.id} checkpointed to {writer.path}.',
         )
       except BaseException as e:  # pylint: disable=broad-except
         experiment.error(
-            f'Failed to save example {example.id} to {writer.path}. '
+            f'Failed to checkpoint example {example.id} to {writer.path}. '
             f'Error: {e}, Stacktrace: \n{traceback.format_exc()}.',
         )
         raise e
@@ -316,7 +350,7 @@ class SequenceWriter:
   def __init__(self, path: str):
     self._lock = threading.Lock()
     self._path = path
-    self._sequence_writer = pg.io.open_sequence(path, 'w')
+    self._sequence_writer = pg.io.open_sequence(path, 'a')
 
   @property
   def path(self) -> str:

langfun/core/eval/v2/checkpointing_test.py

@@ -52,10 +52,20 @@ class SequenceWriterTest(unittest.TestCase):
       self.assertEqual(len(list(iter(f))), 1)
 
 
-class PerExampleCheckpointerTest(unittest.TestCase):
+class CheckpointerTest(unittest.TestCase):
+
+  def assert_found_in_log(self, experiment, message):
+    found_error_log = False
+    for log_entry in experiment._log_entries:
+      if log_entry.message.startswith(message):
+        found_error_log = True
+        break
+    self.assertTrue(found_error_log)
+
+
+class PerExampleCheckpointerTest(CheckpointerTest):
 
   def test_checkpointing(self):
-    pg.defaults.loggers.use_stdout()
     root_dir = os.path.join(tempfile.gettempdir(), 'per_example_checkpointer')
     experiment = eval_test_helper.test_experiment()
     checkpoint_filename = 'checkpoint.jsonl'
@@ -85,8 +95,90 @@ class PerExampleCheckpointerTest(unittest.TestCase):
     for leaf in experiment.leaf_nodes:
       self.assertEqual(leaf.progress.num_skipped, num_processed[leaf.id])
 
+    # Test warm start without reprocess.
+    root_dir = os.path.join(tempfile.gettempdir(), 'per_example_checkpointer2')
+    experiment = eval_test_helper.test_experiment()
+    _ = experiment.run(
+        root_dir, 'new', runner='sequential', plugins=[checkpointer],
+        warm_start_from=run.output_root
+    )
+    for leaf in experiment.leaf_nodes:
+      self.assertEqual(leaf.progress.num_skipped, num_processed[leaf.id])
+
+    # Test warm start with reprocess.
+    root_dir = os.path.join(tempfile.gettempdir(), 'per_example_checkpointer3')
+    experiment = eval_test_helper.test_experiment()
+    _ = experiment.run(
+        root_dir, 'new', runner='sequential', plugins=[checkpointer],
+        warm_start_from=run.output_root,
+        reprocess=True
+    )
+    for leaf in experiment.leaf_nodes:
+      self.assertEqual(leaf.progress.num_skipped, 0)
 
-class BulkCheckpointerTest(unittest.TestCase):
+    root_dir = os.path.join(tempfile.gettempdir(), 'per_example_checkpointer4')
+    experiment = eval_test_helper.test_experiment()
+    _ = experiment.run(
+        root_dir, 'new', runner='sequential', plugins=[checkpointer],
+        warm_start_from=run.output_root,
+        reprocess=[1, 2, 3]
+    )
+    for leaf in experiment.leaf_nodes:
+      self.assertEqual(leaf.progress.num_skipped, num_processed[leaf.id] - 3)
+
+  def test_loading_corrupted_checkpoint(self):
+    root_dir = os.path.join(
+        tempfile.gettempdir(),
+        'per_example_checkpointer_with_corrupted_checkpoint'
+    )
+    experiment = eval_test_helper.TestEvaluation()
+    checkpoint_filename = 'checkpoint.jsonl'
+    checkpointer = checkpointing.PerExampleCheckpointer(checkpoint_filename)
+    run = experiment.run(
+        root_dir, 'new', runner='sequential', plugins=[checkpointer]
+    )
+    num_processed = {}
+    for i in range(experiment.num_examples):
+      example = experiment.state.get(i + 1)
+      ckpt = run.output_path_for(experiment, f'checkpoint_{example.id}.jsonl')
+      if not example.has_error:
+        self.assertTrue(pg.io.path_exists(ckpt))
+        with pg.io.open_sequence(ckpt) as f:
+          self.assertEqual(len(list(iter(f))), 1)
+
+        # Simulate corrupting the first checkpoint.
+        if i == 0:
+          pg.io.writefile(ckpt, 'bad file')
+        num_processed[example.id] = i + 1
+
+    root_dir = os.path.join(
+        tempfile.gettempdir(),
+        'per_example_checkpointer_with_corrupted_checkpoint_warm_start'
+    )
+    experiment = eval_test_helper.TestEvaluation()
+    _ = experiment.run(
+        root_dir, 'new', runner='sequential', plugins=[checkpointer],
+        warm_start_from=run.output_root,
+    )
+    for leaf in experiment.leaf_nodes:
+      self.assertEqual(leaf.progress.num_skipped, len(num_processed) - 1)
+    self.assert_found_in_log(experiment, 'Failed to load checkpoint')
+
+  def test_checkpointing_error(self):
+    root_dir = os.path.join(
+        tempfile.gettempdir(),
+        'per_example_checkpointer_with_checkpointing_error'
+    )
+    experiment = (eval_test_helper
+                  .test_experiment_with_example_checkpointing_error())
+    checkpointer = checkpointing.PerExampleCheckpointer('checkpoint.jsonl')
+    _ = experiment.run(
+        root_dir, 'new', runner='parallel', plugins=[checkpointer]
+    )
+    self.assert_found_in_log(experiment, 'Failed to checkpoint')
+
+
+class BulkCheckpointerTest(CheckpointerTest):
 
   def test_checkpointing(self):
     root_dir = os.path.join(tempfile.gettempdir(), 'test_bulk_checkpointer')
@@ -118,6 +210,19 @@ class BulkCheckpointerTest(unittest.TestCase):
     for leaf in experiment.leaf_nodes:
       self.assertEqual(leaf.progress.num_skipped, num_processed[leaf.id])
 
+  def test_checkpointing_error(self):
+    root_dir = os.path.join(
+        tempfile.gettempdir(),
+        'bulk_checkpointer_with_checkpointing_error'
+    )
+    experiment = (eval_test_helper
+                  .test_experiment_with_example_checkpointing_error())
+    checkpointer = checkpointing.BulkCheckpointer('checkpoint.jsonl')
+    _ = experiment.run(
+        root_dir, 'new', runner='parallel', plugins=[checkpointer]
+    )
+    self.assert_found_in_log(experiment, 'Failed to checkpoint')
+
 
 if __name__ == '__main__':
   unittest.main()

langfun/core/eval/v2/eval_test_helper.py

@@ -72,9 +72,65 @@ class TestEvaluation(Evaluation):
     )
 
 
+class BadJsonConvertible(pg.Object):
+
+  def to_json(self, *args, **kwargs):
+    raise ValueError('Cannot convert to JSON.')
+
+
+class TestEvaluationWithExampleCheckpointingError(TestEvaluation):
+  """Test evaluation class with bad example checkpointing."""
+  inputs = test_inputs()
+  metrics = [metrics_lib.Match()]
+
+  def process(self, v):
+    return 1, dict(
+        x=BadJsonConvertible()
+    )
+
+
+class BadHtmlConvertible(pg.Object, pg.views.HtmlTreeView.Extension):
+
+  def _html_tree_view(self, *args, **kwargs):
+    raise ValueError('Cannot render HTML.')
+
+
+class TestEvaluationWithExampleHtmlGenerationError(Evaluation):
+  """Test evaluation class with bad example HTML generation."""
+  inputs = test_inputs()
+  metrics = [metrics_lib.Match()]
+
+  def process(self, v):
+    return 1, dict(
+        x=BadHtmlConvertible()
+    )
+
+
+class TestEvaluationWithIndexHtmlGenerationError(TestEvaluation):
+  """Test evaluation class with bad index HTML generation."""
+
+  def _html_tree_view(self, *args, **kwargs):
+    raise ValueError('Cannot render HTML.')
+
+
 def test_experiment():
   """Returns a test experiment."""
   return Suite([
       TestEvaluation(lm=TestLLM(offset=0)),
       TestEvaluation(lm=TestLLM(offset=pg.oneof(range(5)))),
   ])
+
+
+def test_experiment_with_example_checkpointing_error():
+  """Returns a test experiment with example checkpointing error."""
+  return TestEvaluationWithExampleCheckpointingError()
+
+
+def test_experiment_with_example_html_generation_error():
+  """Returns a test experiment with bad example HTML."""
+  return TestEvaluationWithExampleHtmlGenerationError()
+
+
+def test_experiment_with_index_html_generation_error():
+  """Returns a test experiment with bad index HTML."""
+  return TestEvaluationWithIndexHtmlGenerationError()

langfun/core/eval/v2/evaluation.py

@@ -264,11 +264,21 @@ class Evaluation(experiment_lib.Experiment):
     return self._state
 
   def load_state(
-      self, state_file: str, raise_if_not_exist: bool = False
+      self,
+      state_file: str,
+      *,
+      load_example_metadata: bool = True,
+      filter: Callable[[example_lib.Example], bool] | None = None,  # pylint: disable=redefined-builtin
+      raise_if_not_exist: bool = False
   ) -> None:
     """Loads saved state from a sequence IO file."""
     if pg.io.path_exists(state_file):
-      self._state.load(state_file, self.example_input_by_id)
+      self._state.load(
+          state_file,
+          example_input_by_id=self.example_input_by_id,
+          load_example_metadata=load_example_metadata,
+          filter=filter,
+      )
     elif raise_if_not_exist:
       raise ValueError(f'State file {state_file} does not exist.')
 
@@ -680,14 +690,25 @@ class EvaluationState:
     self._evaluated_examples: dict[int, example_lib.Example] = {}
 
   def load(
-      self, state_file: str, example_input_by_id: Callable[[int], Any]) -> None:
+      self,
+      state_file: str,
+      *,
+      example_input_by_id: Callable[[int], Any] | None = None,
+      load_example_metadata: bool | Callable[
+          [example_lib.Example], bool] = True,
+      filter: Callable[[example_lib.Example], bool] | None = None,  # pylint: disable=redefined-builtin
+  ) -> None:
     """Loads the state from the example sequence file."""
     with pg.io.sequence.open_sequence(state_file) as f:
       for record in f:
         example = pg.from_json_str(
-            record, example_input_by_id=example_input_by_id
+            record,
+            example_input_by_id=example_input_by_id,
+            load_example_metadata=load_example_metadata
         )
         assert isinstance(example, example_lib.Example), example
+        if filter is not None and not filter(example):
+          continue
         self._evaluated_examples[example.id] = example
 
   @property

langfun/core/eval/v2/evaluation_test.py

@@ -138,6 +138,17 @@ class EvaluationTest(unittest.TestCase):
     self.assertEqual(example.usage_summary.uncached.total.total_tokens, 0)
     self.assertEqual(example.usage_summary.uncached.total.num_requests, 0)
 
+    # Test load_state with filter.
+    exp.reset()
+    self.assertEqual(len(exp._state.evaluated_examples), 0)
+    exp.load_state(state_file, filter=lambda x: x.id == 3)
+    self.assertEqual(len(exp._state.evaluated_examples), 1)
+
+    exp.reset()
+    self.assertEqual(len(exp._state.evaluated_examples), 0)
+    exp.load_state(state_file, filter=lambda x: x.id == 1)
+    self.assertEqual(len(exp._state.evaluated_examples), 0)
+
   def test_html_view(self):
     exp = eval_test_helper.TestEvaluation()
     exp.debug('debug message')

langfun/core/eval/v2/example.py

@@ -101,6 +101,7 @@ class Example(pg.JSONConvertible, pg.views.HtmlTreeView.Extension):
       json_value: dict[str, Any],
       *,
       example_input_by_id: Callable[[int], Any] | None = None,
+      load_example_metadata: bool | Callable[['Example'], bool] = False,
       **kwargs
   ) -> 'Example':
     """Creates an example from the JSON representation."""
@@ -128,12 +129,21 @@ class Example(pg.JSONConvertible, pg.views.HtmlTreeView.Extension):
       pg.traverse(example, _visit)
       return list(referred_types)
 
+    # We delay loading the metadata until the other parts of the example are
+    # loaded. So we could apply the filter to decide whether to load the
+    # metadata.
+    metadata_dict = json_value.pop('metadata', None)
     with pg.JSONConvertible.load_types_for_deserialization(
         *example_class_defs(example_input)
     ):
-      return cls(
+      example = cls(
           **{k: pg.from_json(v, **kwargs) for k, v in json_value.items()}
       )
+      if callable(load_example_metadata):
+        load_example_metadata = load_example_metadata(example)
+      if load_example_metadata:
+        example.metadata = pg.from_json(metadata_dict, **kwargs)
+      return example
 
   #
   # HTML rendering.

langfun/core/eval/v2/example_test.py

@@ -70,17 +70,31 @@ class ExampleTest(unittest.TestCase):
     self.assertEqual(
         pg.from_json_str(
             json_str,
-            example_input_by_id=lambda i: inputs[i - 1]
+            example_input_by_id=lambda i: inputs[i - 1],
+            load_example_metadata=True,
         ),
         ex
     )
+    self.assertEqual(
+        pg.from_json_str(
+            json_str,
+            example_input_by_id=lambda i: inputs[i - 1],
+            load_example_metadata=False,
+        ),
+        Example(
+            id=1,
+            input=inputs[0],
+            output=inputs[0].a(1),
+            metadata={}
+        )
+    )
     pg.JSONConvertible._TYPE_REGISTRY._type_to_cls_map.pop(
         inputs[0].a.__type_name__
     )
     pg.JSONConvertible._TYPE_REGISTRY._type_to_cls_map.pop(
         inputs[0].b.__type_name__
     )
-    v = pg.from_json_str(json_str, auto_dict=True)
+    v = pg.from_json_str(json_str, auto_dict=True, load_example_metadata=True)
     v.output.pop('type_name')
     v.metadata.b.pop('type_name')
     self.assertEqual(