evalscope 0.10.0__py3-none-any.whl → 1.2.0__py3-none-any.whl

evalscope/api/benchmark/adapters/image_edit_adapter.py

@@ -0,0 +1,82 @@
+import os
+from typing import Optional
+
+from evalscope.constants import EvalType, FileConstants
+from evalscope.utils import get_logger
+from evalscope.utils.function_utils import thread_safe
+from evalscope.utils.io_utils import jsonl_to_list
+from .text2image_adapter import Text2ImageAdapter
+
+logger = get_logger()
+
+
+class ImageEditAdapter(Text2ImageAdapter):
+    """
+    Support two methods:
+    1. Inference using modelscope pipeline
+    2. Load local inference jsonl file with key to corresponding prompt
+    """
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+        self.local_file = self.extra_params.get('local_file', None)
+        self.id_key = self.extra_params.get('id_key', FileConstants.ID)
+        self.image_key = self.extra_params.get('image_key', FileConstants.IMAGE_PATH)
+        self.local_data = self.load_local_file()
+
+    def load_local_file(self) -> Optional[dict]:
+        if not self.local_file:
+            return None
+
+        # Load file and check
+        data_list = jsonl_to_list(self.local_file)
+        data_dict = {}
+        for record in data_list:
+            if self.image_key not in record:
+                raise ValueError(f"Image key '{self.image_key}' not found in record: {record}, file {self.local_file}")
+            if self.id_key not in record:
+                raise ValueError(f"ID key '{self.id_key}' not found in record: {record}, file {self.local_file}")
+
+            image_path = record[self.image_key]
+            if not os.path.isabs(image_path):
+                image_path = os.path.join(os.path.dirname(self.local_file), image_path)
+            if not os.path.exists(image_path):
+                raise FileNotFoundError(f"Image file '{image_path}' not found.")
+
+            data_dict[record[self.id_key]] = record
+        return data_dict
+
+    def get_image_path_from_id(self, image_id) -> Optional[str]:
+        if not self.local_file:
+            return None
+
+        record = self.local_data.get(image_id)
+        if not record:
+            return None
+
+        return record[self.image_key]
+
+    def _post_process_samples(self):
+        super()._post_process_samples()
+
+        # Add local image path if exists
+        for subset in self.test_dataset.keys():
+            for sample in self.test_dataset[subset]:
+                local_image_path = self.get_image_path_from_id(sample.metadata.get(FileConstants.ID))
+                if local_image_path:
+                    sample.metadata[FileConstants.IMAGE_PATH] = local_image_path
+
+    def sample_filter(self, sample) -> bool:
+        """
+        Filter samples based on metadata availability.
+        If local file is not available, all samples are considered valid.
+        Otherwise, only samples with valid metadata and image path are kept.
+        """
+        if not self.local_data:
+            return True
+        else:
+            sample_id = sample.metadata.get(FileConstants.ID)
+            if (not sample_id) or (not self.get_image_path_from_id(sample_id)):
+                return False
+            return True

evalscope/api/benchmark/adapters/multi_choice_adapter.py

@@ -0,0 +1,86 @@
+from evalscope.api.dataset.dataset import Sample
+from evalscope.api.evaluator import Choices, Target, TaskState
+from evalscope.utils.multi_choices import (
+    FEW_SHOT_TEMPLATE,
+    MultipleChoiceTemplate,
+    format_example,
+    parse_answers,
+    parse_answers_zh,
+    prompt,
+    valid_template,
+)
+from .default_data_adapter import DefaultDataAdapter
+
+
+class MultiChoiceAdapter(DefaultDataAdapter):
+    """
+    Adapter for multi-choice benchmarks.
+    This adapter formats the input for multi-choice questions and handles few-shot examples.
+    """
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+        self.multiple_correct: bool = False
+        """Whether the benchmark allows multiple correct answers."""
+
+    def format_prompt_template(self, sample: Sample) -> str:
+        """
+        Format the basic prompt template with the sample data.
+
+        Args:
+            sample (Sample): The sample object containing the prompt data
+
+        Returns:
+            str: The formatted prompt ready for model input
+        """
+        assert valid_template(self.prompt_template), 'Prompt template is not valid'
+
+        return prompt(
+            question=sample.input,
+            choices=Choices(sample.choices),
+            template=self.prompt_template,
+        )
+
+    def format_fewshot_template(self, fewshot: str, sample: Sample) -> str:
+        """
+        Format the few-shot template with demonstrations and the main prompt.
+
+        Args:
+            fewshot (str): The formatted few-shot demonstration examples
+            sample (Sample): The sample object containing the prompt data
+
+        Returns:
+            str: The complete formatted input with few-shot context
+        """
+
+        few_shot_prompt_template = self.few_shot_prompt_template or (FEW_SHOT_TEMPLATE + self.prompt_template)
+
+        assert valid_template(few_shot_prompt_template), 'Few-shot prompt template is not valid'
+
+        return prompt(
+            question=sample.input, choices=Choices(sample.choices), template=few_shot_prompt_template, fewshot=fewshot
+        )
+
+    def sample_to_fewshot(self, sample: Sample) -> str:
+        """
+        Convert a sample to a few-shot formatted string.
+
+        Args:
+            sample (Sample): The sample object to format
+
+        Returns:
+            str: The formatted few-shot example string
+        """
+        return format_example(question=sample.input, choices=Choices(sample.choices), answer=Target(sample.target))
+
+    def extract_answer(self, prediction: str, task_state: TaskState) -> str:
+        if self.prompt_template in [
+            MultipleChoiceTemplate.CHINESE_SINGLE_ANSWER_TEMPLATE_COT,
+            MultipleChoiceTemplate.CHINESE_SINGLE_ANSWER_TEMPLATE
+        ]:
+            # For Chinese COT template, we use a different extraction method
+            answers = parse_answers_zh(task_state, multiple_correct=self.multiple_correct)
+        else:
+            answers = parse_answers(task_state, multiple_correct=self.multiple_correct)
+        return ''.join(sorted(list(answers)))

evalscope/api/benchmark/adapters/ner_adapter.py

@@ -0,0 +1,212 @@
+from typing import Any, Dict, List, Set, Tuple
+
+from evalscope.api.dataset import Sample
+from evalscope.api.metric.scorer import AggScore, SampleScore, Score
+from evalscope.utils.import_utils import check_import
+from evalscope.utils.logger import get_logger
+from evalscope.utils.ner import (
+    DEFAULT_TAG_FIX_PATTERNS,
+    calculate_bio_metrics,
+    clean_prediction,
+    create_target_text,
+    extract_entities_from_text,
+    extract_spans_from_bio,
+    xml_to_bio_tags,
+)
+from .default_data_adapter import DefaultDataAdapter
+
+logger = get_logger()
+
+
+class NERAdapter(DefaultDataAdapter):
+    """
+    Base adapter class for Named Entity Recognition (NER) tasks.
+
+    This adapter handles converting between BIO tagging schemes and XML-style entity markup,
+    and provides evaluation metrics using seqeval.
+
+    Subclasses should define their entity types and register the benchmark.
+    """
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        # Define mapping from BIO tags to user-friendly tag names
+        self.entity_type_map = {}
+        # Add descriptions for each entity type
+        self.entity_descriptions = {}
+
+        # These will be initialized in setup_entity_mappings
+        self.reverse_entity_map = {}
+        self.entity_list = []
+        self.entities_description = ''
+
+        # Define common error patterns to handle
+        self.tag_fix_patterns = DEFAULT_TAG_FIX_PATTERNS
+
+        check_import('seqeval', 'seqeval', raise_error=True, feature_name='NER metrics')
+        # Note: setup_entity_mappings() should be called by subclasses
+        # after they define their entity_type_map and entity_descriptions
+
+    def setup_entity_mappings(self):
+        """
+        Setup entity mappings and descriptions for prompt formatting.
+        This should be called after entity_type_map and entity_descriptions are defined.
+        """
+        # Reverse mapping for converting back from prediction to evaluation
+        self.reverse_entity_map = {v.lower(): k for k, v in self.entity_type_map.items()}
+
+        # Create list of tags for prompt formatting
+        self.entity_list = [f'<{ent.lower()}>' for ent in self.entity_type_map.values()]
+
+        # Create description of entities for prompt
+        self.entities_description = ', '.join([
+            f'{self.entity_type_map[tag]} ({self.entity_descriptions[tag]})' for tag in self.entity_type_map
+        ])
+
+    def record_to_sample(self, record: Dict[str, Any]) -> Sample:
+        """
+        Convert a record with tokens and NER tags into a Sample.
+        Creates both the raw text input and annotated text target.
+        """
+        tokens: List[str] = record['tokens']
+        ner_tags: List[str] = record['ner_tags']
+
+        # Create the input text by joining tokens
+        input_text = ' '.join(tokens)
+
+        # Process tokens and tags to create annotated target text
+        target_text = create_target_text(tokens, ner_tags, self.entity_type_map)
+
+        # Store tokens and tags in metadata for evaluation
+        metadata = {'tokens': tokens, 'ner_tags': ner_tags}
+
+        return Sample(input=input_text, target=target_text, metadata=metadata)
+
+    def format_prompt_template(self, sample):
+        """
+        Format the prompt with entity types, available tags, and text to annotate.
+        """
+        return self.prompt_template.format(
+            entities=self.entities_description, entity_list=', '.join(self.entity_list), text=sample.input
+        )
+
+    def format_fewshot_template(self, fewshot, sample):
+        """
+        Format the few-shot prompt with all required parameters.
+        """
+        return self.few_shot_prompt_template.format(
+            fewshot=fewshot,
+            entities=self.entities_description,
+            entity_list=', '.join(self.entity_list),
+            text=sample.input
+        )
+
+    def sample_to_fewshot(self, sample: Sample) -> str:
+        """
+        Format a sample as a few-shot example showing original and annotated text.
+        """
+        if not sample.metadata:
+            return ''
+
+        # Format few-shot examples to match the expected response format
+        return f'Input:\n{sample.input}\n\nOutput:\n{sample.target}'
+
+    def match_score(self, original_prediction, filtered_prediction, reference, task_state) -> Score:
+        """
+        Evaluate named entity recognition performance using seqeval.
+        """
+        from seqeval.metrics import accuracy_score, f1_score, precision_score, recall_score
+
+        score = Score(
+            extracted_prediction=filtered_prediction,
+            prediction=original_prediction,
+        )
+
+        try:
+            # Get the original tokens and tags from the reference metadata
+            original_tokens = task_state.metadata['tokens']
+            original_tags = task_state.metadata['ner_tags']
+
+            if not original_tokens or len(original_tokens) == 0:
+                if hasattr(reference, 'metadata') and reference.metadata:
+                    original_tokens = reference.metadata['tokens']
+                    original_tags = reference.metadata['ner_tags']
+
+            # Clean and normalize the prediction
+            cleaned_prediction = clean_prediction(filtered_prediction, self.tag_fix_patterns)
+
+            # Convert XML-style prediction back to BIO tags aligned with original tokens
+            pred_bio_tags = xml_to_bio_tags(cleaned_prediction, original_tokens, self.reverse_entity_map)
+
+            # Use seqeval to calculate metrics
+            # Note: seqeval expects lists of lists (one per sequence)
+            y_true = [original_tags]
+            y_pred = [pred_bio_tags]
+
+            precision = precision_score(y_true, y_pred)
+            recall = recall_score(y_true, y_pred)
+            f1 = f1_score(y_true, y_pred)
+            accuracy = accuracy_score(y_true, y_pred)
+
+            score.value = {'precision': precision, 'recall': recall, 'f1_score': f1, 'accuracy': accuracy}
+
+            # Store tags for aggregation (proper micro-averaging in aggregate_scores)
+            # This way aggregate_scores can compute metrics across all samples at once,
+            # which gives you true micro-averaged scores rather than averaged macro scores.
+            score.metadata = {'y_true': original_tags, 'y_pred': pred_bio_tags}
+        except Exception as e:
+            logger.warning(f'Error evaluating NER prediction: {str(e)}')
+            score.value = {'precision': 0.0, 'recall': 0.0, 'f1_score': 0.0, 'accuracy': 0.0}
+
+        return score
+
+    def aggregate_scores(self, sample_scores: List[SampleScore]) -> List[AggScore]:
+        """
+        Aggregate metrics across all samples using seqeval.
+        """
+        from seqeval.metrics import accuracy_score, f1_score, precision_score, recall_score
+
+        # Collect all predictions and references
+        y_true_all = []
+        y_pred_all = []
+
+        for ss in sample_scores:
+            # Extract the BIO tags from metadata if available
+            # You may need to store these during match_score
+            if hasattr(ss.score, 'metadata') and 'y_true' in ss.score.metadata and 'y_pred' in ss.score.metadata:
+                y_true_all.append(ss.score.metadata['y_true'])
+                y_pred_all.append(ss.score.metadata['y_pred'])
+
+        if not y_true_all:
+            # Fallback: calculate averages from individual scores
+            num_samples = len(sample_scores)
+            avg_precision = sum(ss.score.value.get('precision', 0.0) for ss in sample_scores) / num_samples
+            avg_recall = sum(ss.score.value.get('recall', 0.0) for ss in sample_scores) / num_samples
+            avg_f1 = sum(ss.score.value.get('f1_score', 0.0) for ss in sample_scores) / num_samples
+            avg_accuracy = sum(ss.score.value.get('accuracy', 0.0) for ss in sample_scores) / num_samples
+        else:
+            # Use seqeval for micro-averaged metrics across all samples
+            avg_precision = precision_score(y_true_all, y_pred_all)
+            avg_recall = recall_score(y_true_all, y_pred_all)
+            avg_f1 = f1_score(y_true_all, y_pred_all)
+            avg_accuracy = accuracy_score(y_true_all, y_pred_all)
+
+        num_samples = len(sample_scores)
+
+        agg_scores = [
+            AggScore(
+                metric_name='precision',
+                score=avg_precision,
+                num=num_samples,
+                metadata={'type': 'seqeval-micro-average'}
+            ),
+            AggScore(
+                metric_name='recall', score=avg_recall, num=num_samples, metadata={'type': 'seqeval-micro-average'}
+            ),
+            AggScore(metric_name='f1_score', score=avg_f1, num=num_samples, metadata={'type': 'seqeval-micro-average'}),
+            AggScore(
+                metric_name='accuracy', score=avg_accuracy, num=num_samples, metadata={'type': 'seqeval-accuracy'}
+            )
+        ]
+
+        return agg_scores

evalscope/api/benchmark/adapters/text2image_adapter.py

@@ -0,0 +1,157 @@
+import base64
+import os
+
+from evalscope.api.dataset import Sample
+from evalscope.api.evaluator import TaskState
+from evalscope.api.messages.chat_message import ChatMessageUser
+from evalscope.api.messages.content import ContentImage
+from evalscope.api.metric import Score
+from evalscope.api.model import ChatCompletionChoice, Model, ModelOutput
+from evalscope.api.registry import get_metric
+from evalscope.constants import EvalType, FileConstants
+from evalscope.utils import get_logger
+from evalscope.utils.function_utils import thread_safe
+from .default_data_adapter import DefaultDataAdapter
+
+logger = get_logger()
+
+
+class Text2ImageAdapter(DefaultDataAdapter):
+    """Text to Image Adapter for benchmarks."""
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+        self.add_aggregation_name = False  # Do not add aggregation name in the report by default
+
+    def load_from_disk(self, **kwargs):
+        return super().load_from_disk(use_local_loader=True)
+
+    def record_to_sample(self, record) -> Sample:
+        """Convert a record dictionary to a Sample object."""
+        return Sample(
+            input=[ChatMessageUser(content=record['prompt'])],
+            metadata={
+                'prompt': record['prompt'],
+                'category': record.get('category', ''),
+                'tags': record.get('tags', []),
+                FileConstants.ID: record[FileConstants.ID],
+                FileConstants.IMAGE_PATH: record.get(FileConstants.IMAGE_PATH,
+                                                     ''),  # Optional field for existing image path
+            }
+        )
+
+    def _on_inference(self, model: Model, sample: Sample) -> ModelOutput:
+        """
+        Hook method called during the actual inference process.
+
+        This method executes the model inference and can be overridden
+        to implement custom inference logic or model interaction patterns.
+
+        Args:
+            model (Model): The model to use for inference
+            sample (Sample): The sample to process
+
+        Returns:
+            ModelOutput: The raw output from the model
+        """
+        if self.eval_type == EvalType.MOCK_LLM:
+            return ModelOutput(
+                model=model.name,
+                choices=[ChatCompletionChoice.from_content('')],
+            )
+        else:
+            # Execute model inference with the processed input and any tools
+            model_output = model.generate(input=sample.input, tools=sample.tools)
+            return model_output
+
+    def _on_inference_end(
+        self, model: Model, sample: Sample, model_output: ModelOutput, output_dir: str, **kwargs
+    ) -> TaskState:
+        """
+        Hook method called after inference completes. Save generated images to output_dir.
+
+        Args:
+            model (Model): The model that performed inference
+            sample (Sample): The processed sample
+            model_output (ModelOutput): The raw model output
+            output_dir (str): The directory where the model output was saved
+
+        Returns:
+            TaskState: Complete state object for the inference task
+        """
+        if self.eval_type == EvalType.MOCK_LLM:
+            return TaskState(
+                model=model.name,
+                sample=sample,
+                messages=[model_output.message],
+                output=model_output,
+                completed=True,
+            )
+        else:
+            image_id = f'{sample.metadata.get(FileConstants.ID, sample.id)}_{sample.group_id}'
+            output_path = os.path.join(output_dir, 'images', f'{image_id}.png')
+            if not os.path.exists(os.path.dirname(output_path)):
+                os.makedirs(os.path.dirname(output_path))
+            # get base64 image from model_output
+            content = model_output.message.content[0]
+
+            assert isinstance(content, ContentImage), 'Expected ContentImage in model output'
+
+            image_base64 = content.image
+            with open(output_path, 'wb') as f:
+                f.write(base64.b64decode(image_base64))
+
+            sample.metadata[FileConstants.IMAGE_PATH] = output_path
+            return TaskState(
+                model=model.name,
+                sample=sample,
+                messages=[model_output.message],
+                output=model_output,
+                completed=True,
+            )
+
+    # NOTE: thread safe is needed, since we can't batch inference here.
+    @thread_safe
+    def match_score(
+        self, original_prediction: str, filtered_prediction: str, reference: str, task_state: TaskState
+    ) -> Score:
+        # Get prediction and prompt from task state
+        image_path = task_state.metadata.get(FileConstants.IMAGE_PATH, original_prediction)
+        prompt = task_state.input[0].content
+        meta = task_state.metadata
+
+        # Initialize the score object with prediction details
+        score = Score(
+            extracted_prediction=image_path,
+            prediction=image_path,
+        )
+
+        # Calculate scores for each configured metric
+        for metric in self.metric_list:
+            try:
+                if isinstance(metric, str):
+                    metric_name = metric
+                    metric_scorer = get_metric(metric)  # Get metric implementation from registry
+                    metric_func = metric_scorer()  # Instantiate the metric scorer
+                elif isinstance(metric, dict):
+                    metric_name = list(metric.keys())[0]
+                    metric_cls = get_metric(metric_name)
+                    metric_func = metric_cls(**metric[metric_name])  # Initialize with parameters
+                metric_score = metric_func(image_path, prompt)[0]
+
+                # fine-granular metrics
+                category = meta.get('category')
+                if category:
+                    metric_name = f'{metric_name}_{category}'
+                if isinstance(metric_score, dict):
+                    for k, v in metric_score.items():
+                        score.value[f'{metric_name}_{k}'] = v.cpu().item()
+                else:
+                    score.value[metric_name] = metric_score.cpu().item()
+            except Exception as e:
+                logger.error(f'Error calculating metric {metric}: {e}')
+                score.value[metric_name] = 0
+                score.metadata[metric_name] = f'error: {str(e)}'
+
+        return score

evalscope/api/benchmark/adapters/vision_language_adapter.py

@@ -0,0 +1,8 @@
+from .default_data_adapter import DefaultDataAdapter
+
+
+class VisionLanguageAdapter(DefaultDataAdapter):
+    """Adapter for vision-language benchmarks. e.g., image captioning, visual question answering, etc."""
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)