langfun 0.1.2.dev202509120804__py3-none-any.whl → 0.1.2.dev202512040805__py3-none-any.whl

langfun/core/eval/v2/experiment.py

@@ -139,10 +139,10 @@ class Experiment(lf.Component, pg.views.HtmlTreeView.Extension):
 
   # Checkpointing
 
-  Experiments support checkpointing, which is enabled by default. It allows 
+  Experiments support checkpointing, which is enabled by default. It allows
   users to resume their experiments from a saved state. When an experiment runs,
-  it creates a new directory for that run and saves the current state to a
-  checkpoint file. If the experiment is interrupted or fails, users can resume
+  it creates a new directory for that run and saves its progress to checkpoint
+  files. If the experiment is interrupted or fails, users can resume
   it by specifying the 'id' or 'warm_start_from' argument (shown above) to
   seamlessly continue from previously saved state without starting over.
 
@@ -169,7 +169,7 @@ class Experiment(lf.Component, pg.views.HtmlTreeView.Extension):
 
   # Experiment Plugins
 
-  Experiment can be extended by plugins. Plugins can listen to the events of
+  Experiments can be extended by plugins. Plugins can listen to the events of
   experiment execution and produce additional outputs. For example, a plugin
   can be added to an experiment to generate additional metrics or to save
   additional data to a database. More details will be added in the future.
@@ -657,7 +657,30 @@ class Experiment(lf.Component, pg.views.HtmlTreeView.Extension):
 
 @pg.use_init_args(['children'])
 class Suite(Experiment):
-  """A suite of evaluations."""
+  """A suite of evaluations.
+
+  `lf.eval.Suite` groups multiple `lf.eval.Evaluation` or other `Suite`
+  objects into a single experiment, allowing them to be run, managed, and
+  reported together.
+
+  **Example:**
+
+  ```python
+  import langfun as lf
+
+  suite = lf.eval.Suite([
+      MyEval(lm=lf.llms.Gpt4()),
+      MyEval(lm=lf.llms.Gemini()),
+      lf.eval.Suite([
+          AnotherEval(lm=lf.llms.Gpt4()),
+          AnotherEval(lm=lf.llms.Gemini())
+      ])
+  ])
+
+  # Run all evaluations in the suite
+  run_info = suite.run('/path/to/my/suite_run')
+  ```
+  """
 
   children: Annotated[
       list[Experiment], 'A list of child experiments.'
@@ -791,7 +814,14 @@ class RunId(pg.Object):
 
 
 class Run(pg.Object, pg.views.html.HtmlTreeView.Extension):
-  """A run of an experiment."""
+  """Represents a single run of an experiment.
+
+  A `Run` object holds all the configurations for executing an experiment,
+  such as the experiment definition, input/output directories, and flags
+  controlling the execution behavior (e.g., error handling, checkpointing).
+  It also provides utility methods for accessing run-specific paths and
+  filtering examples for evaluation.
+  """
 
   root_dir: Annotated[
       str,
@@ -818,10 +848,10 @@ class Run(pg.Object, pg.views.html.HtmlTreeView.Extension):
   ] = None
 
   example_ids: Annotated[
-      list[int] | None,
+      list[int] | Callable[[Experiment], list[int]] | None,
       (
-          'The example IDs to run. If None, it will run all examples. '
-          'Though '
+          'The example IDs to run. Or a callable for determining the examples '
+          'to run based on the experiment. If None, it will run all examples. '
       )
   ] = None
 
@@ -937,10 +967,13 @@ class Run(pg.Object, pg.views.html.HtmlTreeView.Extension):
     """Returns the example IDs to evaluate."""
     if not experiment.is_leaf:
       return set()
-    return set(
-        self.example_ids if self.example_ids else
-        range(1, experiment.num_examples + 1)
-    )
+    if self.example_ids is None:
+      return set(range(1, experiment.num_examples + 1))
+    elif isinstance(self.example_ids, Callable):
+      return set(self.example_ids(experiment))
+    else:
+      assert isinstance(self.example_ids, list), self.example_ids
+      return set(self.example_ids)
 
   def examples_to_reprocess(self, experiment: Experiment) -> set[int]:
     """Returns the example IDs to reprocess per request."""
@@ -971,7 +1004,13 @@ class Run(pg.Object, pg.views.html.HtmlTreeView.Extension):
 
 
 class Runner(pg.Object):
-  """Interface for experiment runner."""
+  """Interface for experiment runner.
+
+  A runner is responsible for executing the evaluations within an experiment
+  based on the configuration specified in a `Run` object. Different runners
+  can implement different execution strategies, such as sequential or parallel
+  processing of examples and evaluations.
+  """
 
   # Class-level variable for registering the runner.
   NAME = None
@@ -1010,7 +1049,37 @@ class Runner(pg.Object):
 
 
 class Plugin(lf.Component):
-  """Base class for experiment plugins."""
+  """Base class for experiment plugins.
+
+  Plugins provide a mechanism to extend the behavior of an experiment run
+  by hooking into various events during the lifecycle of experiment and
+  example execution, such as `on_run_start`, `on_experiment_complete`,
+  `on_example_start`, etc. They can be used for custom logging, monitoring,
+  or result processing.
+  """
+
+  @classmethod
+  def is_per_example(cls) -> bool:
+    """Returns whether the plugin is per example only.
+
+    Per-example plugins can be installed on individual workers when examples
+    are evaluated by multiple processes in parallel.
+    """
+
+    def same_code(method1, method2):
+      return method1.__code__ == method2.__code__
+    return all(
+        same_code(method1, method2)
+        for method1, method2 in [
+            (Plugin.on_run_start, cls.on_run_start),
+            (Plugin.on_run_complete, cls.on_run_complete),
+            (Plugin.on_run_abort, cls.on_run_abort),
+            (Plugin.on_experiment_start, cls.on_experiment_start),
+            (Plugin.on_experiment_skipped, cls.on_experiment_skipped),
+            (Plugin.on_experiment_complete, cls.on_experiment_complete),
+            (Plugin.on_experiment_abort, cls.on_experiment_abort),
+        ]
+    )
 
   def on_run_start(
       self,

langfun/core/eval/v2/experiment_test.py

@@ -433,5 +433,24 @@ class RunnerTest(unittest.TestCase):
           pass
 
 
+class PluginTest(unittest.TestCase):
+
+  def test_per_example_only(self):
+
+    class PerExamplePlugin(experiment_lib.Plugin):
+
+      def on_example_complete(self, runner, experiment, example):
+        print('on_example_complete')
+
+    self.assertTrue(PerExamplePlugin.is_per_example())
+
+    class NonPerExamplePlugin(experiment_lib.Plugin):
+
+      def on_experiment_complete(self, runner, experiment):
+        print('on_example_complete')
+
+    self.assertFalse(NonPerExamplePlugin.is_per_example())
+
+
 if __name__ == '__main__':
   unittest.main()

langfun/core/eval/v2/metric_values.py

@@ -20,7 +20,15 @@ import pyglove as pg
 
 
 class MetricValue(pg.Object):
-  """Base class for metric values."""
+  """Base class for metric values.
+
+  `MetricValue` is the base class for representing aggregated metric values
+  in an evaluation. It accumulates data points from individual examples,
+  each consisting of a value and an optional weight, associated with an example
+  ID. Subclasses must implement `reduce` method to compute a single float value
+  from accumulated data points, and `scalar_repr` to provide a string
+  representation of the reduced value.
+  """
 
   class DataPoint(pg.Object):
     """A data point for a metric value."""
@@ -88,6 +96,14 @@ class MetricValue(pg.Object):
         self.increment_total()
     return self
 
+  def merge_from(self, other: 'MetricValue') -> 'MetricValue':
+    """Merges the values from another metric value."""
+    self._weighted_sum += other._weighted_sum  # pylint: disable=protected-access
+    with pg.notify_on_change(False), pg.allow_writable_accessors(True):
+      self.data_points.extend(other.data_points)
+      self.increment_total(other.total)
+    return self
+
   def __gt__(self, other: Union['MetricValue', float]) -> bool:
     if isinstance(other, self.__class__):
       return float(self) > float(other)
@@ -133,7 +149,13 @@ class MetricValue(pg.Object):
 
 
 class Rate(MetricValue):
-  """Representing a rate in range [0, 1]."""
+  """Metric value representing a rate in range [0, 1].
+
+  `Rate` is used for metrics that compute a rate, such as accuracy or error
+  rate. The final value is computed as the weighted sum of accumulated values
+  divided by the total number of examples. It's displayed as a percentage
+  (e.g., 90.0%).
+  """
 
   def reduce(self) -> float:
     return self._weighted_sum / self.total
@@ -145,7 +167,13 @@ class Rate(MetricValue):
 
 
 class Average(MetricValue):
-  """Average of a aggregated values."""
+  """Metric value representing an average of accumulated values.
+
+  `Average` is used for metrics that compute an average score across examples
+  (e.g., average quality score). The final value is computed as the weighted
+  sum of accumulated values divided by the number of data points.
+  It's displayed as a float with 3 decimal places (e.g., 4.750).
+  """
 
   def reduce(self) -> float:
     if not self.data_points:

langfun/core/eval/v2/metric_values_test.py

@@ -51,6 +51,22 @@ class RateTest(unittest.TestCase):
     self.assertEqual(rate.total, 0)
     self.assertTrue(math.isnan(float(rate)))
 
+  def test_merge_from(self):
+    rate1 = metric_values.Rate()
+    rate1.add(1, 1.0, 1.0, increment_total=True)
+    rate2 = metric_values.Rate()
+    rate2.add(2, 0.0, 1.0, increment_total=True)
+    rate1.merge_from(rate2)
+    self.assertEqual(rate1.total, 2)
+    self.assertEqual(float(rate1), 0.5)
+    self.assertEqual(
+        rate1.data_points,
+        [
+            metric_values.MetricValue.DataPoint(1, 1.0, 1.0),
+            metric_values.MetricValue.DataPoint(2, 0.0, 1.0),
+        ],
+    )
+
 
 class AverageTest(unittest.TestCase):
 
@@ -75,6 +91,22 @@ class AverageTest(unittest.TestCase):
     average.reset()
     self.assertEqual(average.total, 0)
 
+  def test_merge_from(self):
+    avg1 = metric_values.Average()
+    avg1.add(1, 1.0, 0.5, increment_total=True)
+    avg2 = metric_values.Average()
+    avg2.add(2, 0.0, 1.0, increment_total=True)
+    avg1.merge_from(avg2)
+    self.assertEqual(avg1.total, 2)
+    self.assertEqual(float(avg1), 0.25)
+    self.assertEqual(
+        avg1.data_points,
+        [
+            metric_values.MetricValue.DataPoint(1, 1.0, 0.5),
+            metric_values.MetricValue.DataPoint(2, 0.0, 1.0),
+        ],
+    )
+
 
 if __name__ == '__main__':
   unittest.main()

langfun/core/eval/v2/metrics.py

@@ -29,7 +29,15 @@ Average = metric_values.Average
 
 
 class Metric(pg.Object, pg.views.HtmlTreeView.Extension):
-  """Interface for an evaluation metric."""
+  """Interface for an evaluation metric.
+
+  A metric is used to evaluate the quality of the outputs produced by an
+  evaluation. It works by auditing each processed example via its `audit`
+  method, which in turn calls the user-overridable `_audit` method to perform
+  metric-specific logic and update metric values. Metrics can compute multiple
+  values (e.g., precision, recall, F1 score) which are exposed via the
+  `values` method.
+  """
 
   name: Annotated[
       str,
@@ -44,24 +52,43 @@ class Metric(pg.Object, pg.views.HtmlTreeView.Extension):
     self._label_group = None
     self._lock = threading.Lock()
 
-  def audit(self, example: example_lib.Example) -> dict[str, Any]:
-    """Audits a processed example and returns metric metadata for it."""
-    # NOTE(daiyip): the metric values are being updated concurrently, so we
-    # uses a lock to avoid race condition. We might consider relaxing the lock
-    # later if metric auditing becomes a bottleneck.
-    with self._lock:
-      for v in self.values():
-        v.increment_total()
+  def update(
+      self,
+      example: example_lib.Example,
+      force_recompute: bool = False
+  ) -> dict[str, Any]:
+    """Updates metric values with a processed example.
 
-      metadata = self._audit(example)
+    Args:
+      example: The processed example.
+      force_recompute: Whether to force recompute the metric metadata even if
+        they are already present.
 
-      self._update_view()
-      return metadata
+    Returns:
+      A dict of metric metadata.
+    """
+    if (force_recompute
+        or example.metric_metadata is None
+        or self.name not in example.metric_metadata):
+      metadata = self.compute_metric_metadata(example)
+    else:
+      metadata = example.metric_metadata[self.name]
+    self.update_metric_values(example.id, metadata)
+    self._update_view()
+    return metadata
 
   @abc.abstractmethod
-  def _audit(self, example: example_lib.Example) -> dict[str, Any]:
+  def compute_metric_metadata(
+      self, example: example_lib.Example
+  ) -> dict[str, Any]:
     """Subclasses should override this method to implement the metric logic."""
 
+  @abc.abstractmethod
+  def update_metric_values(
+      self, example_id: int, metric_metadata: dict[str, Any]
+  ) -> None:
+    """Update metric values based on metric metadata."""
+
   @abc.abstractmethod
   def values(self) -> list[metric_values.MetricValue]:
     """Returns all the values computed by this metric."""
@@ -71,6 +98,12 @@ class Metric(pg.Object, pg.views.HtmlTreeView.Extension):
     for v in self.values():
       v.reset()
 
+  def merge_from(self, other: 'Metric') -> 'Metric':
+    """Merges the values from another metric."""
+    for v1, v2 in zip(self.values(), other.values()):
+      v1.merge_from(v2)
+    return self
+
   def _update_view(self):
     """Refreshes the metric values."""
     if self._label_group is None:
@@ -169,7 +202,15 @@ class Metric(pg.Object, pg.views.HtmlTreeView.Extension):
 
 
 class MetricBase(Metric):
-  """Base class for common metrics."""
+  """Base class for common metrics.
+
+  `MetricBase` provides common functionalities for metrics, such as automatic
+  error counting based on whether an example has an error during evaluation.
+  It distinguishes between Object-Oriented Programming (OOP) errors
+  (e.g. `MappingError` during structured output generation) and other errors.
+  Subclasses should implement `_audit_processed` for metric computation on
+  successfully processed examples.
+  """
 
   oop_errors: Rate | None = Rate()
   non_oop_errors: Rate | None = Rate()
@@ -183,27 +224,67 @@ class MetricBase(Metric):
     super().reset()
     self._error_breakdown = collections.defaultdict(list)
 
-  def _audit(self, example: example_lib.Example) -> dict[str, Any]:
-    """Audits the evaluation example after processing."""
+  def compute_metric_metadata(
+      self, example: example_lib.Example
+  ) -> dict[str, Any]:
+    """Computes the metric metadata for the example."""
     if example.error is None:
-      return self._audit_processed(example)
+      return self._compute_metric_metadata(example)
+    return self._compute_metric_metadata_with_processing_error(example)
+
+  def update_metric_values(
+      self,
+      example_id: int,
+      metric_metadata: dict[str, Any]
+  ) -> None:
+    """Collects the metric metadata."""
+    # NOTE(daiyip): the metric values are being updated concurrently, so we
+    # uses a lock to avoid race condition. We might consider relaxing the lock
+    # later if metric auditing becomes a bottleneck.
+    with self._lock:
+      for v in self.values():
+        v.increment_total()
+
+    if 'error' in metric_metadata:
+      self._update_metric_values_with_processing_error(
+          example_id, metric_metadata
+      )
     else:
-      return self._audit_error(example)
+      self._update_metric_values(example_id, metric_metadata)
+
+  @abc.abstractmethod
+  def _compute_metric_metadata(
+      self,
+      example: example_lib.Example
+  ) -> dict[str, Any]:
+    """Computes the metric metadata for the example."""
 
-  def _audit_error(self, example: example_lib.Example) -> dict[str, Any]:
+  def _compute_metric_metadata_with_processing_error(
+      self,
+      example: example_lib.Example
+  ) -> dict[str, Any]:
     """Audits the evaluation example after processing."""
     assert example.error is not None
-    tag = example.error.tag
-    if tag.startswith('MappingError'):
-      self.oop_errors.add(example.id, 1)
-    else:
-      self.non_oop_errors.add(example.id, 1)
-    self._error_breakdown[tag].append(example.id)
-    return dict(error=tag)
+    return dict(error=example.error.tag)
 
   @abc.abstractmethod
-  def _audit_processed(self, example: example_lib.Example) -> dict[str, Any]:
-    """Audits the evaluation example after processing."""
+  def _update_metric_values(self, metadata: dict[str, Any]) -> None:
+    """Update metric values based metric metadata."""
+
+  def _update_metric_values_with_processing_error(
+      self,
+      example_id: int,
+      metric_metadata: dict[str, Any]
+  ) -> None:
+    """Updates metric values with processing error."""
+    error_tag = metric_metadata.get('error')
+    assert error_tag is not None, (example_id, metric_metadata)
+    self._error_breakdown[error_tag].append(example_id)
+    if error_tag.startswith('MappingError'):
+      self.oop_errors.add(example_id, 1)
+    else:
+      self.non_oop_errors.add(example_id, 1)
+    self._error_breakdown[error_tag].append(example_id)
 
   def _oop_errors_breakdown(self) -> str | None:
     """Returns the OOP error breakdown as a string."""
@@ -229,7 +310,13 @@ class MetricBase(Metric):
 
 
 class Match(MetricBase):
-  """Metric for matching outputs against groundtruth."""
+  """Metric for matching outputs against ground truth.
+
+  This metric computes match and mismatch rates by comparing the output of
+  an example with its ground truth. By default, it looks for a `groundtruth`
+  attribute in `example.input` for comparison. Users can customize this behavior
+  by subclassing `Match` and overriding the `match` method.
+  """
 
   name = 'match'
   matches: Rate = Rate()
@@ -257,20 +344,30 @@ class Match(MetricBase):
       )
     return pg.eq(output, groundtruth)
 
-  def _audit_processed(self, example: example_lib.Example) -> dict[str, Any]:
-    """Audits the evaluation example after processing."""
+  def _compute_metric_metadata(
+      self, example: example_lib.Example
+  ) -> dict[str, Any]:
+    """Computes the metric metadata for the example."""
     metadata = {}
-    is_match = self.match(example.input, example.output)
-    if isinstance(is_match, tuple):
-      is_match, metadata = is_match
-    if is_match:
-      self.matches.add(example.id, 1)
-      metadata['match'] = True
-    else:
-      self.mismatches.add(example.id, 1)
-      metadata['mismatch'] = True
+    is_correct = self.match(example.input, example.output)
+    if isinstance(is_correct, tuple):
+      is_correct, metadata = is_correct
+
+    metadata['is_correct'] = is_correct
     return metadata
 
+  def _update_metric_values(
+      self, example_id: int, metadata: dict[str, Any]
+  ) -> None:
+    """Update metric values based metric metadata."""
+    is_correct = metadata.get('is_correct')
+    assert is_correct is not None, (example_id, metadata)
+    if is_correct:
+      self.matches.add(example_id, 1)
+    else:
+      assert not is_correct
+      self.mismatches.add(example_id, 1)
+
   def values(self) -> list[metric_values.MetricValue]:
     """Returns all the values computed by this metric."""
     return [
@@ -302,7 +399,14 @@ class Match(MetricBase):
 
 
 class Score(MetricBase):
-  """Base class for scoring."""
+  """Base class for scoring metrics.
+
+  `Score` is a base class for metrics that assign a numerical score to each
+  example's output (e.g., evaluating quality on a scale of 1-5).
+  It automatically computes the average score across all examples.
+  Subclasses must implement the `score` method to define how an example
+  should be scored.
+  """
 
   name = 'score'
   average_score: Average = Average()
@@ -322,16 +426,25 @@ class Score(MetricBase):
       A float score. Or a tuple of (score, metadata).
     """
 
-  def _audit_processed(self, example: example_lib.Example) -> dict[str, Any]:
-    """Audits the evaluation example after processing."""
+  def _compute_metric_metadata(
+      self, example: example_lib.Example
+  ) -> dict[str, Any]:
+    """Computes the metric metadata for the example."""
     metadata = {}
     score = self.score(example.input, example.output)
     if isinstance(score, tuple):
       score, metadata = score
-    self.average_score.add(example.id, score)
     metadata['score'] = score
     return metadata
 
+  def _update_metric_values(
+      self, example_id: int, metadata: dict[str, Any]
+  ) -> None:
+    """Update metric values based metric metadata."""
+    score = metadata.get('score')
+    assert score is not None, (example_id, metadata)
+    self.average_score.add(example_id, score)
+
   def values(self) -> list[metric_values.MetricValue]:
     """Returns all the values computed by this metric."""
     return [