langfun 0.0.2.dev20240429__py3-none-any.whl → 0.1.2.dev202501140804__py3-none-any.whl

langfun/core/eval/base.py

@@ -24,10 +24,9 @@ import os
 import re
 import threading
 import time
+import types
 from typing import Annotated, Any, Callable, Iterator, Literal, Optional, Sequence, Type, Union
 
-from absl import app
-from absl import flags
 import langfun.core as lf
 import langfun.core.coding as lf_coding
 from langfun.core.llms.cache import in_memory
@@ -40,7 +39,8 @@ class Evaluable(lf.Component):
 
   EXPERIMENT_JSON = 'experiment.json'
   RESULT_JSON = 'result.json'
-  FAILURES_JSON = 'failures.json'
+  OOP_FAILURES_JSON = 'oop_failures.json'
+  NON_OOP_FAILURES_JSON = 'non_oop_failures.json'
   INDEX_HTML = 'index.html'
   SUMMARY_HTML = 'summary.html'
 
@@ -215,6 +215,7 @@ class Evaluable(lf.Component):
       summary: bool = True,
       pivot_field: str = 'lm',
       from_root: bool = True,
+      timeout: int | None = None,
       **kwargs,
   ) -> Union['Summary', pg.Dict]:
     """Run the evaluation, which fills and returns the result."""
@@ -242,7 +243,7 @@ class Evaluable(lf.Component):
       ):
         if show_progress:
           lf.concurrent.ProgressBar.update(
-              progress_bar, postfix='LOADING SAVED RESULTS...', color='yellow'
+              progress_bar, status='LOADING SAVED RESULTS...', color='yellow'
           )
         if self.try_load_result():
           run_status = 'CACHED'
@@ -265,13 +266,14 @@ class Evaluable(lf.Component):
             verbose=verbose,
             progress_bar=progress_bar,
             label=label,
+            timeout=timeout,
             **kwargs,
         )
 
         if should_save:
           if show_progress:
             lf.concurrent.ProgressBar.update(
-                progress_bar, postfix='SAVING RESULTS...', color='yellow'
+                progress_bar, status='SAVING RESULTS...', color='yellow'
             )
 
           # Save evaluation results.
@@ -284,7 +286,7 @@ class Evaluable(lf.Component):
       if show_progress:
         lf.concurrent.ProgressBar.update(
             progress_bar,
-            postfix=self._completion_status(run_status),
+            status=self._completion_status(run_status),
             color='green',
         )
     else:
@@ -340,7 +342,7 @@ class Evaluable(lf.Component):
               f'[#{leaf.index} - {leaf.node.id}]',
               total=leaf.node.num_examples if leaf.enabled else 0,
               color='cyan' if leaf.enabled else 'yellow',
-              postfix=None if leaf.enabled else 'SKIPPED.')
+              status=None if leaf.enabled else 'SKIPPED.')
 
         # Run leaf groups in parallel.
         try:
@@ -354,17 +356,17 @@ class Evaluable(lf.Component):
           # Save results for non-leaf nodes.
           lf.concurrent.ProgressBar.update(
               overview_bar,
-              postfix='SAVING RESULTS...',
+              status='SAVING RESULTS...',
               color='yellow')
 
           for node in self.nonleaf_nodes:
-            node._result = {c.id: c.result for c in node.children}  # pylint: disable=protected-access
+            node._result = {c.id: c.result for c in node.leaf_nodes}  # pylint: disable=protected-access
             if should_save:
               node.save(result=False, report=False)
 
           if should_save and summary:
             lf.concurrent.ProgressBar.update(
-                overview_bar, postfix='FINALIZING SUMMARY...'
+                overview_bar, status='FINALIZING SUMMARY...'
             )
 
             summary.save(os.path.join(self.root_dir, Evaluable.SUMMARY_HTML))
@@ -378,7 +380,7 @@ class Evaluable(lf.Component):
           # Signal all task completed by making the bar green.
           lf.concurrent.ProgressBar.update(
               overview_bar,
-              postfix='COMPLETED',
+              status='COMPLETED',
               color='green')
 
         finally:
@@ -398,6 +400,7 @@ class Evaluable(lf.Component):
       verbose: bool,
       progress_bar: int | None,
       label: str | None,
+      timeout: int | None = None,
       **kwargs,
   ) -> None:
     """Run the evaluate and fill `self.result`. Subclass to implement."""
@@ -528,37 +531,14 @@ class Evaluable(lf.Component):
     self._render_message(self.dryrun_output, s)
 
   def _render_message(self, message: lf.Message, s: io.StringIO) -> None:
-    for m in message.trace():
-      if 'lm-input' in m.tags:
-        text_color = 'green'
-      elif 'lm-response' in m.tags:
-        text_color = 'blue'
-      else:
-        text_color = 'black'
-
-      s.write(
-          f'<div style="color: {text_color}; white-space: pre-wrap;'
-          'padding: 10px; border: 1px solid; margin-top: 10px">'
-      )
-      s.write(m.get('formatted_text', m.text))
-      if m.result is not None:
-        s.write(
-            '<div style="color: magenta; white-space: pre-wrap;'
-            'padding: 10px; border: 1px solid; margin: 10px">'
-        )
-        s.write(pg.format(m.result))
-        s.write('</div>')
-      if 'usage' in m.metadata:
-        s.write(
-            '<div style="background-color: #EEEEEE; color: black; '
-            'white-space: pre-wrap; padding: 10px; border: 0px solid; '
-            'margin: 10px">'
-            f'prompt: {m.usage.prompt_tokens} tokens, '
-            f'response: {m.usage.completion_tokens} tokens, '
-            f'total: {m.usage.total_tokens} tokens'
-            '</div>'
+    s.write(
+        message.to_html_str(
+            extra_flags=dict(
+                include_message_metadata=False,
+                source_tag=['lm-input', 'lm-response'],
+            )
         )
-      s.write('</div>')
+    )
 
   @classmethod
   def from_dir(
@@ -598,7 +578,6 @@ class _LeafNode:
 @pg.use_init_args(['children'])
 class Suite(Evaluable):
   """Evaluation suite."""
-
   children: Annotated[list[Evaluable], 'Child evaluation sets or suites.']
 
   # Use empty ID as suite is just a container of child evaluations.
@@ -753,10 +732,12 @@ class Evaluation(Evaluable):
 
   # Constants.
   CACHE_JSON = 'cache.json'
-  FAILURES_HTML = 'failures.html'
+  OOP_FAILURES_HTML = 'oop_failures.html'
+  NON_OOP_FAILURES_HTML = 'non_oop_failures.html'
 
   @functools.cached_property
   def hash(self) -> str:
+    """Returns the semantic-based hash of the evaluation."""
     if self.is_deterministic:
       identity = pg.format(self._identifiers(), compact=True)
     else:
@@ -805,6 +786,10 @@ class Evaluation(Evaluable):
     """Returns the complete rate."""
     return self.num_completed / self.num_examples
 
+  #
+  # Properties on failures.
+  #
+
   @property
   def failures(self) -> list[tuple[Any, Exception]]:
     """Returns the failed examples and their errors."""
@@ -815,6 +800,15 @@ class Evaluation(Evaluable):
     """Returns the number of failed examples."""
     return len(self.failures)
 
+  @functools.cached_property
+  def failure_breakdown(self) -> dict[str, int]:
+    """Returns the breakdown of failures."""
+    breakdown = collections.defaultdict(int)
+    for _, error in self.failures:
+      breakdown[_error_key(error)] += 1
+    sorted_items = sorted(breakdown.items(), key=lambda x: x[1], reverse=True)
+    return pg.Dict({x[0]: x[1] for x in sorted_items})
+
   @property
   def failure_rate(self) -> float:
     """Returns the failure rate in range [0, 1]."""
@@ -822,6 +816,46 @@ class Evaluation(Evaluable):
       return 0.0
     return self.num_failures / self.num_completed
 
+  @functools.cached_property
+  def oop_failures(self) -> list[tuple[Any, lf_structured.MappingError]]:
+    """Returns the OOP failures."""
+    return [item for item in self.failures
+            if isinstance(item[1], lf_structured.MappingError)]
+
+  @property
+  def num_oop_failures(self) -> int:
+    """Returns the number of OOP failures."""
+    return len(self.oop_failures)
+
+  @property
+  def oop_failure_rate(self) -> float:
+    """Returns the OOP failure rate in range [0, 1]."""
+    if self.num_completed == 0:
+      return 0.0
+    return self.num_oop_failures / self.num_completed
+
+  @functools.cached_property
+  def non_oop_failures(self) -> list[tuple[Any, Exception]]:
+    """Returns the OOP failures."""
+    return [item for item in self.failures
+            if not isinstance(item[1], lf_structured.MappingError)]
+
+  @property
+  def num_non_oop_failures(self) -> int:
+    """Returns the number of non-OOP failures."""
+    return len(self.non_oop_failures)
+
+  @property
+  def non_oop_failure_rate(self) -> float:
+    """Returns the non-OOP failure rate in range [0, 1]."""
+    if self.num_completed == 0:
+      return 0.0
+    return self.num_non_oop_failures / self.num_completed
+
+  #
+  # Properties on usage.
+  #
+
   @property
   def has_usage(self) -> bool:
     """Returns True if token usage is enabled."""
@@ -905,7 +939,7 @@ class Evaluation(Evaluable):
 
     fields = list(cls.__schema__.values())
     fields.insert(0, (self.completion_prompt_field, pg.typing.Str()))
-    pg.symbolic.update_schema(cls, fields, extend=False)
+    cls.update_schema(fields, extend=False)
 
   def _maybe_adjust_examples_for_completion(
       self,
@@ -976,13 +1010,22 @@ class Evaluation(Evaluable):
     self._total_prompt_tokens = 0
     self._total_completion_tokens = 0
     self._num_usages = 0
+    self.__dict__.pop('oop_failures', None)
+    self.__dict__.pop('non_oop_failures', None)
+
+  @property
+  def oop_failures_link(self) -> str | None:
+    """Returns the link to the OOP failures page."""
+    if self.dir is None:
+      return None
+    return self.link(os.path.join(self.dir, Evaluation.OOP_FAILURES_HTML))
 
   @property
-  def failures_link(self) -> str | None:
-    """Returns the link to the failures page."""
+  def non_oop_failures_link(self) -> str | None:
+    """Returns the link to then non-OOP failures page."""
     if self.dir is None:
       return None
-    return self.link(os.path.join(self.dir, Evaluation.FAILURES_HTML))
+    return self.link(os.path.join(self.dir, Evaluation.NON_OOP_FAILURES_HTML))
 
   def _dryrun(
       self,
@@ -992,11 +1035,11 @@ class Evaluation(Evaluable):
       verbose: bool,
       **kwargs,
   ) -> None:
-    # Set the example for dryrun.
-    example = example or self.examples[0]
-
     # We make a copy to avoid pollute the state of current object.
     copy: Evaluation = self.clone()
+
+    # Set the example for dryrun.
+    example = example or copy.examples[0]
     copy.__dict__['examples'] = [example]
 
     # We set the symbolic parent of the cloned to access contextual information
@@ -1011,23 +1054,36 @@ class Evaluation(Evaluable):
           color='green',
       )
 
-    with lf.use_settings(debug=debug):
-      output_message = copy.process(example, **(self.additional_args or {}))
-      if self.schema is None:
-        output = output_message.text
-      else:
-        output = output_message.result
+    error, output_message = None, None
 
-    if verbose:
+    try:
+      with lf.use_settings(debug=debug):
+        output_message = copy.process(example, **(self.additional_args or {}))
+        self.process_output(example, output_message)
+
+        if self.schema is None:
+          output = output_message.text
+        else:
+          output = output_message.result
+
+      if verbose:
+        lf.console.write('')
+        lf.console.write(
+            str(output),
+            title='OUTPUT',
+            color='blue',
+        )
+    except lf_structured.MappingError as e:
       lf.console.write('')
       lf.console.write(
-          str(output),
-          title='OUTPUT',
-          color='blue',
+          str(e),
+          title='ERROR',
+          color='red',
       )
+      error = e
 
-    copy.audit(example, output_message, None, dryrun=True)
-    result = copy.summarize()
+    copy.audit(1, example, output_message, error, dryrun=True)
+    result = copy.finalize()
 
     if verbose:
       lf.console.write('')
@@ -1048,9 +1104,13 @@ class Evaluation(Evaluable):
       verbose: bool,
       progress_bar: int | None,
       label: str | None,
+      timeout: int | None = None,
       **kwargs,
   ) -> None:
     # Setup examples.
+    # Reset examples so it could be read from the input functor.
+    self.__dict__.pop('examples', None)
+
     if end is None:
       end = len(self.examples)
     examples = self.examples[start:end]
@@ -1059,20 +1119,24 @@ class Evaluation(Evaluable):
     with lf.use_settings(debug=debug, cache=self.cache):
       self._reset()
 
-      def _process(example: Any):
+      def _process(idx_and_example: Any):
         # NOTE(daiyip): set the `input` symbol of the globals to None, so LLM
         # generated code with calls to `input` will raise an error, thus not
         # blocking the evaluation.
+        _, example = idx_and_example
         with lf_coding.context(input=None):
-          return self.process(example, **(self.additional_args or {}))
+          output_message = self.process(example, **(self.additional_args or {}))
+          self.process_output(example, output_message)
+          return output_message
 
       try:
-        for example, message, error in lf.concurrent_map(
+        for (idx, example), message, error in lf.concurrent_map(
             _process,
-            examples,
+            enumerate(examples),
             max_workers=self.max_workers,
             show_progress=progress_bar or False,
             status_fn=self._status,
+            timeout=timeout,
         ):
           if error is not None:
             message = (
@@ -1080,14 +1144,14 @@ class Evaluation(Evaluable):
                 if isinstance(error, lf_structured.MappingError)
                 else None
             )
-          self.audit(example, message, error)
+          self.audit(idx + 1, example, message, error)
       finally:
         # Save cache upon completion or interruption.
         if self.dir and self.cache:
           self.cache.save()
 
     # Summarize result.
-    self._result = self.summarize()
+    self._result = self.finalize()
     if verbose:
       lf.console.write(
           str(self.result),
@@ -1101,7 +1165,7 @@ class Evaluation(Evaluable):
 
   def process(self, example: Any, **kwargs) -> lf.Message:
     """Process an example and returns its output."""
-    prompt = self.prompt.render(example=example).text
+    prompt = lf.Template.from_value(self.prompt, example=example)
     if self.method == 'call':
       return lf_structured.call(
           prompt,
@@ -1129,7 +1193,9 @@ class Evaluation(Evaluable):
     else:
       assert self.method == 'complete', self.method
       assert isinstance(self.schema.spec, pg.typing.Object), self.schema
-      input_value = self.schema.spec.cls.partial(prompt)
+      # TODO(daiyip): Currently multi-modal inputs within the prompt for
+      # completion is not supported.
+      input_value = self.schema.spec.cls.partial(prompt.render().text)
       return lf_structured.complete(
           input_value,
           lm=self.lm,
@@ -1140,16 +1206,48 @@ class Evaluation(Evaluable):
           **kwargs,
       )
 
+  def process_output(self, example: Any, output: lf.Message) -> None:
+    """Process the output for an example.
+
+    Subclasses can override this method to generate and attach additional
+    metadata for debugging purpose. For example, draw bounding boxes on the
+    input image based on LLM predicted boxes and attach to output_message's
+    metadata.
+
+    Example:
+
+      class BoundingBoxEval(lf.eval.Matching):
+        ...
+        def process_output(example, output):
+          output.metadata.image_with_bbox = draw_bboxes(
+              example.image, output.result)
+
+    Args:
+      example: User input.
+      output: LLM's output message. Users could attach additional
+        information to the message, which will be shown in debugging
+    """
+    del example, output
+
   def _status(self, progress: lf.concurrent.Progress) -> dict[str, Any]:
+    status = {'Model': self.lm.model_id}
+    status.update(self._eval_status(progress))
+
+    if progress.last_error is not None:
+      status['LastError'] = progress.last_error_str()
+    if progress.timeit_summary:
+      status['TimeIt'] = progress.timeit_summary_str()
+    return status
+
+  def _eval_status(self, progress: lf.concurrent.Progress) -> dict[str, Any]:
     return {
-        'Model': self.lm.model_id,
-        'Succeeded': f'%.{self.report_precision}f%% (%d/%d)' % (
-            progress.success_rate * 100,
+        'Succeeded': '%s (%d/%d)' % (
+            self._format_rate(progress.success_rate),
             progress.succeeded,
             progress.completed,
         ),
-        'Failed': f'%.{self.report_precision}f%% (%d/%d)' % (
-            progress.failure_rate * 100,
+        'Failed': '%s (%d/%d)' % (
+            self._format_rate(progress.failure_rate),
             progress.failed,
             progress.completed,
         ),
@@ -1159,22 +1257,21 @@ class Evaluation(Evaluable):
     assert self.result is not None
     m = self.result.metrics
     return (
-        f'COMPLETED(%s): Successes=%.{self.report_precision}f%% (%d/%d)'
-        f' Failures=%.{self.report_precision}f%% (%d/%d)'
+        'COMPLETED(%s): Successes=%s(%d/%d) Failures=%s (%d/%d)'
         % (
             run_status,
-            (1 - m.failure_rate) * 100,
+            self._format_rate(1 - m.failure_rate),
             m.total - m.failures,
             m.total,
-            m.failure_rate * 100,
+            self._format_rate(m.failure_rate),
             m.failures,
             m.total,
         )
     )
 
-  def summarize(self) -> pg.Dict:
-    """Summarizes the evaluation result."""
-    if self.cache:
+  def finalize(self) -> pg.Dict:
+    """Finalizes the evaluation result."""
+    if self.cache is not None:
       cache_stats = dict(
           use_cache=True,
           num_queries=self.cache.stats.num_queries,
@@ -1201,7 +1298,7 @@ class Evaluation(Evaluable):
             id=self.id,
             dir=self.dir,
             model=self.lm.model_id,
-            prompt_template=lf.text_formatting.decolored(str(self.prompt)),
+            prompt_template=pg.decolor(str(self.prompt)),
             method=self.method,
             schema_fn=str(self.schema_fn),
         ),
@@ -1210,38 +1307,47 @@ class Evaluation(Evaluable):
             total=self.num_completed,
             failures=self.num_failures,
             failure_rate=self.failure_rate,
+            oop_failures=self.num_oop_failures,
+            oop_failure_rate=self.oop_failure_rate,
+            non_oop_failures=self.num_non_oop_failures,
+            non_oop_failure_rate=self.non_oop_failure_rate,
+            failure_breakdown=self.failure_breakdown,
         ),
         usage=usage,
     )
     return result
 
-  def summarize_html(self) -> str:
+  def summary_card(self) -> str:
+    """Returns summary card in HTML."""
     s = io.StringIO()
     definition = _html_repr(self, compact=False, escape=True)
     s.write('<div><table><tr><td>')
+    self._render_link(
+        s,
+        definition,
+        self.hash,
+        '',
+        lambda: self.link(self.dir),
+    )
     if self.result is None:
       s.write(
-          f'<a target="_blank" title="{definition}" '
-          f'href="{self.link(self.dir)}">{self.hash}</a>'
           '</td></tr><tr><td>'
           '<span style="color: gray">(IN-PROGRESS...)</span>'
       )
     else:
-      s.write(
-          f'<a target="_blank" title="{definition}" '
-          f'href="{self.index_link}">{self.hash}</a>'
-          '</td></tr><tr><td>'
-      )
-      self._render_metric(s)
+      if self.dir:
+        s.write(f' &nbsp;[<a href="{self.link(self.dir)}">dir</a>]')
+      s.write('</td></tr><tr><td>')
+      self._render_summary_metrics(s)
 
       # Summarize average usage.
-      if self.result.usage is not None:
-        self._render_usage(s)
+      if self.result.usage:
+        self._render_summary_usage(s)
 
     s.write('</td></tr></table></div>')
     return s.getvalue()
 
-  def _render_usage(self, s: io.StringIO) -> None:
+  def _render_summary_usage(self, s: io.StringIO) -> None:
     """Renders usage in HTML."""
     usage = self.result.usage
     total = usage.total_prompt_tokens + usage.total_completion_tokens
@@ -1255,22 +1361,79 @@ class Evaluation(Evaluable):
         f'" style="color:gray">({total} tokens)</a>'
     )
 
-  def _render_metric(self, s: io.StringIO) -> None:
+  def _render_link(self,
+                   s: io.StringIO,
+                   title: str,
+                   text: str,
+                   style: str,
+                   url_fn: Callable[[], str]) -> None:
+    """Renders a link in HTML."""
+    s.write(
+        f'<a target="_blank" title="{title}" style="{style}"'
+    )
+    if self.dir:
+      s.write(f' href="{url_fn()}"')
+    s.write(f'>{text}</a>')
+
+  def _render_summary_metrics(self, s: io.StringIO) -> None:
     """Renders metrics in HTML."""
     assert self.result is not None
     m = self.result.metrics
-    s.write(
-        '<a title="Failures (%d/%d)" href="%s" style="color:red">%s</a>'
-        % (
-            m.failures,
-            m.total,
-            self.failures_link,
-            f'%.{self.report_precision}f%% ' % (m.failure_rate * 100),
-        )
+
+    # OOP failures.
+    oop_failure_title = f'OOP failures ({m.oop_failures}/{m.total})'
+    if m.oop_failures:
+      oop_failure_title += '&#013;'
+      for name, count in m.failure_breakdown.items():
+        if name.startswith('MappingError'):
+          oop_failure_title += '&#013;%s: %s (%d/%d)' % (
+              name.removeprefix('MappingError.'),
+              self._format_rate(count / m.total),
+              count,
+              m.total,
+          )
+
+    extra_style = ''
+    if m.oop_failure_rate > 0.1 and m.oop_failures > 3:
+      extra_style = ';font-weight:bold'
+    self._render_link(
+        s,
+        oop_failure_title,
+        self._format_rate(m.oop_failure_rate),
+        f'color:magenta{extra_style}',
+        lambda: self.oop_failures_link,
     )
+    s.write(' | ')
+
+    # Non-OOP failures.
+    non_oop_failure_title = f'Non-OOP failures ({m.non_oop_failures}/{m.total})'
+    if m.non_oop_failures:
+      non_oop_failure_title += '&#013;'
+      for name, count in m.failure_breakdown.items():
+        if not name.startswith('MappingError'):
+          non_oop_failure_title += '&#013;%s: %s (%d/%d)' % (
+              name,
+              self._format_rate(count / m.total),
+              count,
+              m.total,
+          )
+
+    extra_style = ';font-weight:bold' if m.non_oop_failures > 0 else ''
+    self._render_link(
+        s,
+        non_oop_failure_title,
+        self._format_rate(m.non_oop_failure_rate),
+        f'color:red{extra_style}',
+        lambda: self.non_oop_failures_link,
+    )
+
+  def _format_rate(self, rate: float) -> str:
+    """Formats a rate."""
+    return f'%.{self.report_precision}f%% ' % (rate * 100)
 
   def audit(
       self,
+      example_idx: int,
       example: Any,
       message: lf.Message | None,
       error: Exception | None = None,
@@ -1279,6 +1442,7 @@ class Evaluation(Evaluable):
     """Audits the example against the output. Subclasses should override.
 
     Args:
+      example_idx: 1-based index of the example in its dataset.
       example: The input object.
       message: The entire message returned by the LM, which could be used to
         trace the LM input, response and parsed structure. If error is raised
@@ -1287,13 +1451,19 @@ class Evaluation(Evaluable):
       dryrun: Whether or not audition takes place during dryrun.
     """
     if error is not None:
-      self._failures.append((example, str(error)))
+      self._failures.append((example, error))
+
+      # Invalid cache of num_oop_failures.
+      self.__dict__.pop('oop_failures', None)
+      self.__dict__.pop('non_oop_failures', None)
+      self.__dict__.pop('failure_breakdown', None)
+
       if isinstance(error, lf_structured.MappingError):
         message = error.lm_response
     else:
       assert message is not None
       output = message.text if self.schema is None else message.result
-      self.audit_processed(example, output, message, dryrun=dryrun)
+      self.audit_processed(example_idx, example, output, message, dryrun=dryrun)
 
     # Audit usage.
     if message is not None:
@@ -1301,14 +1471,17 @@ class Evaluation(Evaluable):
     self._num_completed += 1
 
   def audit_usage(self, message: lf.Message, dryrun: bool = False) -> None:
+    del dryrun
     for m in message.trace():
-      if 'usage' in m.metadata:
-        self._total_prompt_tokens += m.usage.prompt_tokens
-        self._total_completion_tokens += m.usage.completion_tokens
+      usage = m.metadata.get('usage', None)
+      if usage:
+        self._total_prompt_tokens += usage.prompt_tokens
+        self._total_completion_tokens += usage.completion_tokens
         self._num_usages += 1
 
   def audit_processed(
-      self, example: Any, output: Any, message: lf.Message, dryrun: bool = False
+      self, example_idx: int, example: Any, output: Any, message: lf.Message,
+      dryrun: bool = False
   ) -> None:
     """Audits a successfully processed example. Subclass should override."""
 
@@ -1333,16 +1506,26 @@ class Evaluation(Evaluable):
       # Save failures.
       pg.save(
           [
-              pg.Dict(
-                  input=input, error=lf.text_formatting.decolored(str(error))
-              )
-              for input, error in self.failures
+              pg.Dict(input=input, error=_format_error(error))
+              for input, error in self.oop_failures
           ],
-          os.path.join(self.dir, Evaluation.FAILURES_JSON),
+          os.path.join(self.dir, Evaluation.OOP_FAILURES_JSON),
       )
       pg.save(
-          self._html([self._render_result, self._render_failures]),
-          os.path.join(self.dir, Evaluation.FAILURES_HTML),
+          self._html([self._render_result, self._render_oop_failures]),
+          os.path.join(self.dir, Evaluation.OOP_FAILURES_HTML),
+          file_format='txt',
+      )
+      pg.save(
+          [
+              pg.Dict(input=input, error=_format_error(error))
+              for input, error in self.non_oop_failures
+          ],
+          os.path.join(self.dir, Evaluation.NON_OOP_FAILURES_JSON),
+      )
+      pg.save(
+          self._html([self._render_result, self._render_non_oop_failures]),
+          os.path.join(self.dir, Evaluation.NON_OOP_FAILURES_HTML),
           file_format='txt',
       )
 
@@ -1355,9 +1538,10 @@ class Evaluation(Evaluable):
         '<td>Schema</td>'
         '<td>Additional Args</td>'
     )
-    if self.result.usage is not None:
+    if self.result.usage:
       s.write('<td>Usage</td>')
-    s.write('<td>Failures</td>')
+    s.write('<td>OOP Failures</td>')
+    s.write('<td>Non-OOP Failures</td>')
 
   def _render_result_row(self, s: io.StringIO) -> None:
     s.write(
@@ -1383,18 +1567,31 @@ class Evaluation(Evaluable):
         f'{_html_repr(self.additional_args, compact=False)}</td>'
     )
     # Usage.
-    if self.result.usage is not None:
+    if self.result.usage:
       s.write('<td>')
-      self._render_usage(s)
+      self._render_summary_usage(s)
       s.write('</td>')
 
-    # Failures.
+    # OOP failures.
     s.write(
-        '<td><span style="color:orange">%s</span>%s</td>'
+        '<td><span style="color:magenta">%s</span>%s</td>'
         % (
-            f'%.{self.report_precision}f%%' % (self.failure_rate * 100),
+            self._format_rate(self.oop_failure_rate),
             '<a href="%s">(%d/%d)</a>'
-            % (self.failures_link, self.num_failures, self.num_completed),
+            % (self.oop_failures_link,
+               self.num_oop_failures,
+               self.num_completed),
+        )
+    )
+    # Non-OOP failures.
+    s.write(
+        '<td><span style="color:red">%s</span>%s</td>'
+        % (
+            self._format_rate(self.non_oop_failure_rate),
+            '<a href="%s">(%d/%d)</a>'
+            % (self.non_oop_failures_link,
+               self.num_non_oop_failures,
+               self.num_completed),
         )
     )
 
@@ -1408,31 +1605,99 @@ class Evaluation(Evaluable):
     else:
       return 'cyan'
 
-  def _render_failures(self, s: io.StringIO) -> None:
+  def _render_oop_failures(self, s: io.StringIO) -> None:
+    self._render_failures(s, '^MappingError.*', error_color='magenta')
+
+  def _render_non_oop_failures(self, s: io.StringIO) -> None:
+    self._render_failures(s, '^(?!MappingError).*', error_color='red')
+
+  def _render_failures(
+      self, s: io.StringIO, error_regex: str, error_color: str) -> None:
     """Formats the failed cases into html."""
+    # Failure summary.
     s.write(
-        '<h2> Failed Cases </h2>'
+        '<h2> Error Summary </h2>'
         '<div style="white-space:pre">\n'
         '<table style="border:1px solid">'
-        '<tr class="header"><td>No.</td><td>Input</td><td>Error</td></tr>'
+        '<tr class="header"><td>Error type</td><td>Stats</td></tr>'
+    )
+    error_regex = re.compile(error_regex)
+    if self.result.metrics.failure_breakdown:
+      for name, count in self.result.metrics.failure_breakdown.items():
+        if not error_regex.match(name):
+          continue
+
+        link = f'<a href="#{name}">{name}</a>'
+        error_rate = self._format_rate(count / self.result.metrics.total)
+        stats = (f'<span style="color:{error_color}">{error_rate} '
+                 f'({count}/{self.result.metrics.total})</span>')
+        s.write(f'<tr><td>{link}</td><td>{stats})</td></tr>')
+    s.write(
+        '</table></div>'
+        '<h2> Failed Cases </h2>'
+        '<div style="white-space:pre">'
     )
+    # Failure details by error type.
+    failures_by_error = collections.defaultdict(list)
+    for example, error in self.failures:
+      error_name = _error_key(error)
+      if error_regex.match(error_name):
+        failures_by_error[error_name].append((example, error))
+
+    for error_key, failures in failures_by_error.items():
+      s.write(
+          f'<h3 id="{error_key}"><a href="#{error_key}">{error_key}</a> '
+          f'(count={len(failures)})</h3>'
+          '<table style="border:1px solid">'
+          '<tr class="header"><td>No.</td><td>Input</td>'
+          '<td>LM invocation</td><td>Error</td></tr>'
+      )
+      for i, (example, error) in enumerate(failures):
+        lm_response = None
+        if isinstance(error, lf.structured.MappingError):
+          lm_response = error.lm_response
+          error = error.cause
+
+        bgcolor = 'white' if i % 2 == 0 else '#DDDDDD'
+        s.write(f'<tr style="background-color: {bgcolor}"><td>{i + 1}</td>')
+        s.write('<td style="color:green;white-space:pre-wrap">')
+        s.write(pg.format(example, verbose=False))
+        s.write('</td><td>')
+        if lm_response is not None:
+          self._render_message(lm_response, s)
+        s.write(f'</td><td style="color:{error_color};white-space:pre">')
+        s.write(_format_error(error))
+        s.write('</td></tr>')
+      s.write('</table>')
+    s.write('</div>')
 
-    for i, (example, error) in enumerate(self.failures):
-      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)
-      s.write(f'<td style="color:green;white-space:pre-wrap">{input_str}</td>')
-      error_str = lf.text_formatting.decolored(str(error))
-      s.write(f'<td style="color:red;white-space:pre">{error_str}</td>')
-      s.write('</tr>')
-    s.write('</table></div>')
+  @classmethod
+  def visualize(cls, evaluations: list['Evaluation']) -> str | None:
+    """Visualize the a list of evaluations of this task in HTML."""
+    del evaluations
+    return None
 
 
 @pg.functor()
-def inputs_from(path: str | list[str]) -> list[Any]:
+def inputs_from(path: str | list[str], **kwargs) -> list[Any]:
   """A functor that returns a list of user-defined objects as eval inputs."""
   if isinstance(path, str):
-    return pg.load(path)
+    if path.endswith('.json'):
+      return pg.load(path)
+    elif path.endswith('.jsonl'):
+      return list(iter(pg.open_jsonl(path)))
+    elif path.endswith('.csv'):
+      import pandas as pd  # pylint: disable=g-import-not-at-top
+      dataset_df = pd.read_csv(path, **kwargs)
+      dataset = []
+      for i in range(dataset_df.shape[0]):
+        row = {}
+        for col in dataset_df.columns:
+          row[col] = dataset_df.iloc[i][col]
+        dataset.append(row)
+      return dataset
+    else:
+      raise ValueError(f'Unsupported file format: {path}')
   examples = []
   for p in path:
     examples.extend(pg.load(p))
@@ -1578,7 +1843,7 @@ class Summary(pg.Object):
           if e is None:
             s.write('<span style="color: gray">N/A<span>')
           else:
-            s.write(e.summarize_html())
+            s.write(e.summary_card())
           s.write('</td>')
         s.write('</tr>')
       s.write('</table>')
@@ -1653,13 +1918,22 @@ class Summary(pg.Object):
     s.write('<html><body>')
     for task in sorted(self.tasks(), key=lambda cls: cls.__name__):
       table_id = task.__name__.lower()
+      evaluations = self.select(task=task).evaluations
+      table = Summary.Table.from_evaluations(evaluations, pivot_field)
       s.write('<div>')
-      s.write(f'<a id="{table_id}"')
-      s.write(f'<h2><a href="#{table_id}">{task.__name__}</a></h2>')
-      s.write('</a>')
-      table = Summary.Table.from_evaluations(
-          self.select(task=task).evaluations, pivot_field
+      s.write(
+          f'<a id="{table_id}" href="#{table_id}">'
+          f'<h2>{task.__name__}</h2></a>'
       )
+
+      # Allow users to plugin visualization code (e.g. matplot) in the summary
+      # page.
+      visual_part = task.visualize(evaluations)
+      if visual_part:
+        s.write(visual_part)
+
+      s.write(f'<h4 style="color:gray">{len(evaluations)} experiments</h4>')
+      s.write('<hr/>')
       s.write(table.html())
       s.write('</div>')
     s.write('</body></html>')
@@ -1685,6 +1959,7 @@ class Summary(pg.Object):
                 experiment=entry,
                 dir=entry.dir,
                 metrics=entry.result.metrics if entry.result else None,
+                usage=entry.result.usage if entry.result else None,
             )
         )
       task_results[task.__name__] = results
@@ -1833,6 +2108,20 @@ class Summary(pg.Object):
     return result.join()
 
 
+def _format_error(error: Exception):
+  """Formats an error into a string."""
+  return (f'({error.__class__.__name__}) ' + pg.decolor(str(error)))
+
+
+def _error_key(error: Exception) -> str:
+  """Returns the key for an error."""
+  error_names = []
+  while error is not None:
+    error_names.append(error.__class__.__name__)
+    error = getattr(error, 'cause', None)
+  return '.'.join(error_names)
+
+
 def _html_repr(value: Any, compact: bool = True, escape: bool = False) -> str:
   """Formats prompt in HTML."""
   if type(value) is lf.Template:  # pylint: disable=unidiomatic-typecheck
@@ -1909,41 +2198,200 @@ def monitor_async(
   )
 
 
-def app_run(target: Evaluable):
-  """Runs the target evaluation as an absl app.
+#
+# Named evaluations and experiments support.
+#
 
-  Args:
-    target: An Langfun evaluable object.
-  """
-  flags.DEFINE_string(
-      'root_dir', None, 'Root directory for running the evaluation.'
-  )
 
-  flags.DEFINE_bool(
-      'dryrun', False, 'If True, dryrun the experiment instead of running it.'
-  )
+class _NamedEvaluationRegistry:
+  """Named evaluation registry."""
 
-  flags.DEFINE_bool(
-      'debug', False, 'If True, output prompt and response to the console.'
-  )
+  def __init__(self):
+    self._registry = {}
 
-  flags.DEFINE_bool(
-      'rerun',
-      False,
-      'If True, rerun the experiment even a cached result is found.',
-  )
+  def names(self) -> list[str]:
+    """Returns all registered names."""
+    return sorted(self._registry.keys())
+
+  def get(self, name: str) -> list[Type[Evaluable]]:
+    """Gets an evaluation by name."""
+    matches = []
+    if name in self._registry:
+      matches.append(self._registry[name])
+    else:
+      regex = re.compile(name)
+      for key, cls in self._registry.items():
+        if regex.match(key):
+          matches.append(cls)
+    return matches
+
+  def register(
+      self,
+      name: str,
+      experiment_cls: Type[Evaluable],
+  ):
+    """Register an experiment class."""
+    self._registry[name] = experiment_cls
 
-  FLAGS = flags.FLAGS  # pylint: disable=invalid-name
 
-  def _main(argv):
-    if len(argv) > 1:
-      raise app.UsageError('Too many command-line arguments.')
+_eval_registry = _NamedEvaluationRegistry()
 
-    if FLAGS.root_dir:
-      target.rebind(root_dir=FLAGS.root_dir, raise_on_no_change=False)
-    if FLAGS.dryrun:
-      target.dryrun(debug=FLAGS.debug)
+
+def registered_names() -> list[str]:
+  """Returns all registered names."""
+  return _eval_registry.names()
+
+
+def get_evaluations(evaluation: str | Evaluable) -> list[Evaluable]:
+  """Gets an evaluation experiment by name."""
+  if isinstance(evaluation, str):
+    return [e() for e in _eval_registry.get(evaluation)]
+  return [evaluation]
+
+
+def register(name: str):
+  """Decorator to create a named evaluation class."""
+
+  def _register(func_or_cls: Type[Evaluation] | types.FunctionType):
+    if inspect.isfunction(func_or_cls):
+      e = func_or_cls()
+      if not isinstance(e, Evaluable):
+        raise TypeError(
+            f'The return value of `{func_or_cls}` should be an instance of '
+            '`lf.eval.Evaluable` subclass.'
+        )
+
+      class GeneratedSuite(Suite):
+        # NOTE(daiyip): Delay serialization key registration for generated
+        # class.
+        auto_register = False
+        children = e.children if isinstance(e, Suite) else [e]
+
+      cls = GeneratedSuite
+      cls.__name__ = func_or_cls.__name__
+      cls.__doc__ = func_or_cls.__doc__
+      cls.__qualname__ = func_or_cls.__qualname__
+      cls.__module__ = getattr(func_or_cls, '__module__', 'wrapper')
+      cls.register_for_deserialization(cls.__type_name__)
+
+    elif issubclass(func_or_cls, Evaluable):
+      cls = func_or_cls
     else:
-      target.run(debug=FLAGS.debug, rerun=FLAGS.rerun)
+      raise ValueError(f'Unsupported type: {type(func_or_cls)}')
+
+    _eval_registry.register(name, cls)
+    return cls
+
+  return _register
+
+
+def get(
+    root_dir: str,
+    evaluations: list[str | Evaluable],
+    filter: Union[                    # pylint: disable=redefined-builtin
+        str,                          # Regex to filter evaluation based on ID.
+        Callable[[Evaluable], bool],  # Custom filter function.
+        None                          # No filtering (Default).
+    ] = None,                         # pylint: disable=bad-whitespace
+    patches: list[Union[
+        str,                                    # String-based PyGlove patcher.
+        pg.patching.Patcher,                    # PyGlove patcher object.
+        Callable[[pg.KeyPath, Any, Any], Any],  # PyGlove rebind function.
+    ]] | None = None,                           # pylint: disable=bad-whitespace
+) -> Suite:
+  """Gets a suite from a list of patched evaluations.
+
+  Args:
+    root_dir: The root directory of the experiment.
+    evaluations: A list of evaluations to be included in the suite.
+    filter: A regular expression (str) for selecting sub-experiments of matched
+      IDs, or a filter function to filter the evaluations.
+    patches: A list of patches to be applied to the suite. Each element can be
+      a string (for string-based patcher), a `pg.patching.Patcher` object, or
+      a rebind function (e.g. `pg.rebind`). See `lf.eval.patch_*` for more
+      details.
+
+  Returns:
+    A suite of selected `lf.eval.Evaluation` objects.
+  """
+  matches = []
+  for e in evaluations:
+    matches.extend(get_evaluations(e))
+
+  if not matches:
+    raise ValueError('No evaluations found.')
+
+  suite = Suite(matches, root_dir=root_dir)
+  if patches:
+    suite = pg.patch(suite, patches)
+
+  if isinstance(filter, str):
+    regex = re.compile(filter)
+    filter = lambda x: bool(regex.match(x.id))
+
+  if filter:
+    suite = Suite(
+        [leaf for leaf in suite.leaf_nodes if filter(leaf)], root_dir=root_dir)
+  return suite
+
+
+def run(
+    root_dir: str,
+    evaluations: list[str | Evaluable],
+    filter: Union[                    # pylint: disable=redefined-builtin
+        str,                          # Regex to filter evaluation based on ID.
+        Callable[[Evaluable], bool],  # Custom filter function.
+        None                          # No filtering (Default).
+    ] = None,                         # pylint: disable=bad-whitespace
+    patches: list[Union[
+        str,                                    # String-based PyGlove patcher.
+        pg.patching.Patcher,                    # PyGlove patcher object.
+        Callable[[pg.KeyPath, Any, Any], Any],  # PyGlove rebind function.
+    ]] | None = None,                           # pylint: disable=bad-whitespace
+    mode: Literal['run', 'rerun', 'dryrun', 'noop'] = 'run',
+    debug: bool = False,
+    print_definition: bool = False,
+    **kwargs,
+) -> Suite:
+  """Run selected evaluations with patching.
+
+  Args:
+    root_dir: The root directory of the experiment.
+    evaluations: A list of evaluations to be included in the suite.
+    filter: A regular expression (str) for selecting sub-experiments of matched
+      IDs, or a filter function to filter the evaluations.
+    patches: A list of patches to be applied to the suite. Each element can be
+      a string (for string-based patcher), a `pg.patching.Patcher` object, or
+      a rebind function (e.g. `pg.rebind`). See `lf.eval.patch_*` for more
+      details.
+    mode: The mode to run the suite. "run" to run the suite, with reusing
+      existing results if available; "rerun" to rerun all evaluations even if
+      there are existing results; "dryrun" to dryrun the suite; and "noop"
+      to do nothing.
+    debug: Whether to run in debug mode.
+    print_definition: Whether to print the experiment definition.
+    **kwargs: Additional arguments to be passed to dryrun/run the suite.
+
+  Returns:
+    A suite of selected `lf.eval.Evaluation` objects.
+  """
+  suite = get(root_dir, evaluations, patches=patches, filter=filter)
+  if print_definition:
+    lf.console.write(
+        pg.format(
+            suite,
+            compact=False,
+            verbose=False,
+            hide_default_values=True,
+            python_format=True,
+        ),
+        title='[EXPERIMENT DEFINITION]',
+        color='blue',
+    )
 
-  app.run(_main)
+  if mode == 'run':
+    rerun = mode == 'rerun'
+    suite.run(debug=debug, rerun=rerun, **kwargs)
+  elif mode == 'dryrun':
+    suite.dryrun(debug=debug, **kwargs)
+  return suite