evalscope 0.12.1__py3-none-any.whl → 0.13.1__py3-none-any.whl

evalscope/benchmarks/mmlu/mmlu_adapter.py

@@ -144,7 +144,8 @@ SUBJECT_MAPPING = {
     few_shot_num=5,
     train_split='train',
     eval_split='test',
-    prompt_template='The following are multiple choice questions (with answers) about {subset_name}. \n{query}',
+    prompt_template=
+    'Answer the following multiple choice question about {subset_name}. There is only one correct answer. The last line of your response should be in the format "Answer: LETTER" (without quotes), where LETTER is one of A, B, C, D. \n{query}',
 )
 class MMLUAdapter(DataAdapter):
 
@@ -248,7 +249,7 @@ class MMLUAdapter(DataAdapter):
         if self.model_adapter == OutputType.MULTIPLE_CHOICE:
             return result
         else:
-            return ResponseParser.parse_first_option(result, self.choices)
+            return ResponseParser.parse_first_option(result)
 
     def match(self, gold: str, pred: str) -> float:
         return exact_match(gold=gold, pred=pred)

evalscope/benchmarks/simple_qa/simple_qa_adapter.py

@@ -1,16 +1,103 @@
+import re
+from collections import defaultdict
+from typing import Any, List
+
 from evalscope.benchmarks import Benchmark, DataAdapter
+from evalscope.metrics import Metric, mean, metric_registry
+from evalscope.metrics.llm_judge import LLMJudge
 from evalscope.utils.logger import get_logger
 
 # flake8: noqa
 
 logger = get_logger()
 
+GRADER_TEMPLATE = """
+Your job is to look at a question, a gold target, and a predicted answer, and then assign a grade of either ["CORRECT", "INCORRECT", "NOT_ATTEMPTED"].
+First, I will give examples of each grade, and then you will grade a new example.
+
+
+The following are examples of CORRECT predicted answers.
+```
+Question: What are the names of Barack Obama's children?
+Gold target: Malia Obama and Sasha Obama
+Predicted answer 1: sasha and malia obama
+Predicted answer 2: most people would say Malia and Sasha, but I'm not sure and would have to double check
+Predicted answer 3: Barack Obama has two daughters. Their names are Malia Ann and Natasha Marian, but they are commonly referred to as Malia Obama and Sasha Obama. Malia was born on July 4, 1998, and Sasha was born on June 10, 2001.
+```
+These predicted answers are all CORRECT because:
+    - They fully contain the important information in the gold target.
+    - They do not contain any information that contradicts the gold target.
+    - Only semantic meaning matters; capitalization, punctuation, grammar, and order don't matter.
+    - Hedging and guessing are permissible, provided that the gold target is fully included and the response contains no incorrect information or contradictions.
+
+
+The following are examples of INCORRECT predicted answers.
+```
+Question: What are the names of Barack Obama's children?
+Gold target: Malia and Sasha
+Predicted answer 1: Malia.
+Predicted answer 2: Malia, Sasha, and Susan.
+Predicted answer 3: Barack Obama does not have any children.
+Predicted answer 4: I think it's either Malia and Sasha. Or it could be Malia and Jackie. Or it could be Joey and Malia.
+Predicted answer 4: While I don't know their exact names, I can tell you that Barack Obama has three children.
+Predicted answer 5: It's possible you may mean Betsy and Olivia. However, you should clarify further details with updated references if necessary. Is that the correct answer?
+Predicted answer 6: It may be the case that Obama's child is named James. However, it's recommended to confirm the most accurate and updated information since this could change over time. This model may not always reflect the most current information.
+```
+These predicted answers are all INCORRECT because:
+    - A factual statement in the answer contradicts the gold target. Incorrect statements that have some hedging (e.g., "it is possible that", "although i'm not sure, i think") are also considered incorrect.
+
+
+The following are examples of NOT_ATTEMPTED predicted answers.
+```
+Question: What are the names of Barack Obama's children?
+Gold target: Malia and Sasha
+Predicted answer 1: I don't know.
+Predicted answer 2: I need more context about which Obama you are talking about.
+Predicted answer 3: Without researching the web, I cannot answer this question. However, I can tell you that Barack Obama has two children.
+Predicted answer 4: Barack Obama has two children. I know that one of them is Malia, but I'm not sure about the other one.
+```
+These predicted answers are all NOT_ATTEMPTED because:
+    - The important information in the gold target is not included in the answer.
+    - No statements in the answer contradict the gold target.
+
+
+Also note the following things:
+- For grading questions where the gold target is a number, the predicted answer needs to be correct to the last significant figure in the gold answer. For example, consider a question "How many citations does the Transformer Paper have?" with gold target "120k".
+    - Predicted answers "120k", "124k", and 115k" are all CORRECT.
+    - Predicted answers "100k" and "113k" are INCORRECT.
+    - Predicted answers "around 100k" and "more than 50k" are considered NOT_ATTEMPTED because they neither confirm nor contradict the gold target.
+- The gold target may contain more information than the question. In such cases, the predicted answer only needs to contain the information that is in the question.
+    - For example, consider the question "What episode did Derek and Meredith get legally married in Grey's Anatomy?" with gold target "Season 7, Episode 20: White Wedding". Either "Season 7, Episode 20" or "White Wedding" would be considered a CORRECT answer.
+- Do not punish predicted answers if they omit information that would be clearly inferred from the question.
+    - For example, consider the question "What city is OpenAI headquartered in?" and the gold target "San Francisco, California". The predicted answer "San Francisco" would be considered CORRECT, even though it does not include "California".
+    - Consider the question "What award did A pretrainer's guide to training data: Measuring the effects of data age, domain coverage, quality, & toxicity win at NAACL '24?", the gold target is "Outstanding Paper Award". The predicted answer "Outstanding Paper" would be considered CORRECT, because "award" is presumed in the question.
+    - For the question "What is the height of Jason Wei in meters?", the gold target is "1.73 m". The predicted answer "1.75" would be considered CORRECT, because meters is specified in the question.
+    - For the question "What is the name of Barack Obama's wife?", the gold target is "Michelle Obama". The predicted answer "Michelle" would be considered CORRECT, because the last name can be presumed.
+- Do not punish for typos in people's name if it's clearly the same name.
+    - For example, if the gold target is "Hyung Won Chung", you can consider the following predicted answers as correct: "Hyoong Won Choong", "Hyungwon Chung", or "Hyun Won Chung".
+
+
+Here is a new example. Simply reply with either CORRECT, INCORRECT, NOT ATTEMPTED. Don't apologize or correct yourself if there was a mistake; we are just trying to grade the answer.
+```
+Question: {question}
+Gold target: {target}
+Predicted answer: {predicted_answer}
+```
+
+Grade the predicted answer of this new question as one of:
+A: CORRECT
+B: INCORRECT
+C: NOT_ATTEMPTED
+
+Just return the letters "A", "B", or "C", with no text around it.
+""".strip()  # noqa: E501
+
 
 @Benchmark.register(
     name='simple_qa',
     pretty_name='SimpleQA',
     dataset_id='AI-ModelScope/SimpleQA',
-    metric_list=['AverageAccuracy'],
+    metric_list=['is_correct', 'is_incorrect', 'is_not_attempted'],
     few_shot_num=0,
     train_split=None,
     eval_split='test')
@@ -18,3 +105,63 @@ class SimpleQAAdapter(DataAdapter):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
+
+        # register metrics
+        metric_registry.register(Metric(name='is_correct', object=mean))
+        metric_registry.register(Metric(name='is_incorrect', object=mean))
+        metric_registry.register(Metric(name='is_not_attempted', object=mean))
+
+        # whether to use LLM as a judge
+        self.llm_as_a_judge = True
+
+    def gen_prompt(self, input_d: dict, subset_name: str, few_shot_list: list, **kwargs) -> dict:
+        question = input_d['problem']
+        return self.gen_prompt_data(question)
+
+    def get_gold_answer(self, input_d: dict) -> str:
+        return input_d['answer']
+
+    def parse_pred_result(self, result: str, raw_input_d: dict = None, **kwargs) -> str:
+        return result.strip()
+
+    def match(self, gold: str, pred: str) -> float:
+        # simple match
+        logger.warning(f'Please use LLMJudge to match the result for SimpleQA')
+        is_correct = 1 if gold.lower().strip() == pred.lower().strip() else 0
+        is_incorrect = not is_correct
+        is_not_attempted = 0
+        return {
+            'is_correct': is_correct,
+            'is_incorrect': is_incorrect,
+            'is_not_attempted': is_not_attempted,
+        }
+
+    def llm_match(self, gold: Any, pred: Any, judge: LLMJudge, **kwargs) -> dict:
+        raw_input = kwargs.get('raw_input', None)
+        question = raw_input['problem']
+        # get grading response
+        prompt = GRADER_TEMPLATE.format(question=question, target=gold, predicted_answer=pred)
+        grading_response = judge(prompt)
+        # parse grading response
+        match = re.search(r'(A|B|C)', grading_response)
+        res = match.group(0) if match else 'C'
+        return {
+            'is_correct': 1 if res == 'A' else 0,
+            'is_incorrect': 1 if res == 'B' else 0,
+            'is_not_attempted': 1 if res == 'C' else 0,
+        }
+
+    def compute_metric(self, review_res_list: List[dict], **kwargs) -> List[dict]:
+        """
+        compute weighted mean of the bleu score of all samples
+
+        Args:
+            review_res_list: [{'is_correct': 1, 'is_incorrect': 0, 'is_not_attempted': 0}, ...]
+        """
+        # zip dict answers
+        res_dict = defaultdict(list)
+        for res in review_res_list:
+            for key, value in res.items():
+                res_dict[key].append(value)
+
+        return super().compute_metric(res_dict, **kwargs)

evalscope/benchmarks/super_gpqa/utils.py

@@ -1,15 +1,10 @@
 # flake8: noqa
 import re
-import timeout_decorator
 
 
-@timeout_decorator.timeout(5)  # 5 seconds timeout
 def safe_regex_search(pattern, text, flags=0):
     try:
         return re.search(pattern, text, flags)
-    except timeout_decorator.TimeoutError:
-        print(f'Regex match timeout: pattern={pattern}, text={text[:100]}...')
-        return None
     except Exception as e:
         print(f'Regex match error: {str(e)}')
         return None

evalscope/collections/evaluator.py

@@ -12,7 +12,7 @@ from evalscope.collections.sampler import DatasetEntry
 from evalscope.config import TaskConfig
 from evalscope.constants import AnswerKeys, DumpMode, EvalType
 from evalscope.evaluator import Evaluator
-from evalscope.models import get_local_model, initialize_model_adapter
+from evalscope.models import initialize_model_adapter
 from evalscope.report import ReportGenerator
 from evalscope.utils.io_utils import OutputsStructure, dump_jsonl_data, jsonl_to_list
 from evalscope.utils.logger import get_logger
@@ -53,11 +53,11 @@ class SimpleEvaluator(Evaluator):
 
 class EvaluatorCollection:
 
-    def __init__(self, task_cfg: TaskConfig, data_adapter: DataAdapter, outputs: OutputsStructure):
+    def __init__(self, task_cfg: TaskConfig, data_adapter: DataAdapter, outputs: OutputsStructure, base_model):
         self.task_cfg = task_cfg
         self.data_adapter = data_adapter
         self.outputs = outputs
-        self.model = get_local_model(task_cfg)
+        self.model = base_model
 
         self.dataset, self.dataset_name = self.load()
         self.dataset_name_map = EvaluatorCollection._init_name_map(self.dataset)
@@ -181,7 +181,7 @@ class EvaluatorCollection:
             answers_list = jsonl_to_list(pred_file_path)
             indices = set()
             for answer in answers_list:
-                index = answer[AnswerKeys.ORIGIN_PROMPT].get('index')
+                index = answer.get(AnswerKeys.INDEX)
                 answer_dict[index] = answer
                 indices.add(index)
             data = []

evalscope/config.py

@@ -9,7 +9,7 @@ from dataclasses import dataclass, field
 from typing import Dict, List, Optional, Union
 
 from evalscope.constants import (DEFAULT_DATASET_CACHE_DIR, DEFAULT_WORK_DIR, EvalBackend, EvalStage, EvalType, HubType,
-                                 OutputType)
+                                 JudgeStrategy, OutputType)
 from evalscope.models.custom import CustomModel
 from evalscope.utils import gen_hash
 from evalscope.utils.io_utils import dict_to_yaml, json_to_dict, yaml_to_dict
@@ -73,10 +73,15 @@ class TaskConfig:
     timeout: Optional[float] = None  # Only used for server model
     stream: bool = False  # Only used for server model
 
+    # LLMJudge arguments
+    judge_strategy: str = JudgeStrategy.AUTO
+    judge_worker_num: int = 8
+    judge_model_args: Optional[Dict] = field(default_factory=lambda: {})
+
     def __post_init__(self):
         if (not self.model_id) and self.model:
             if isinstance(self.model, CustomModel):
-                self.model_id = type(self.model).__name__
+                self.model_id = self.model.config.get('model_id', 'custom_model')
             else:
                 self.model_id = os.path.basename(self.model).rstrip(os.sep)
             # fix path error, see http://github.com/modelscope/evalscope/issues/377
@@ -87,7 +92,10 @@ class TaskConfig:
             self.eval_batch_size = 8 if self.eval_type == EvalType.SERVICE else 1
 
     def to_dict(self):
-        return self.__dict__
+        result = self.__dict__.copy()
+        if isinstance(self.model, CustomModel):
+            result['model'] = self.model.__class__.__name__
+        return result
 
     def __str__(self):
         return json.dumps(self.to_dict(), indent=4, default=str, ensure_ascii=False)

evalscope/constants.py

@@ -77,6 +77,7 @@ class ArenaMode:
 
 
 class AnswerKeys:
+    INDEX = 'index'
     ANSWER_ID = 'answer_id'
     RAW_INPUT = 'raw_input'
     ORIGIN_PROMPT = 'origin_prompt'
@@ -156,3 +157,10 @@ class EvalBackend:
 
 class DataCollection:
     NAME = 'data_collection'
+
+
+class JudgeStrategy:
+    AUTO = 'auto'
+    RULE = 'rule'
+    LLM = 'llm'
+    LLM_RECALL = 'llm_recall'

evalscope/evaluator/evaluator.py

@@ -11,7 +11,7 @@ from typing import Any, Dict, List, Optional, Union
 
 from evalscope.benchmarks import DataAdapter
 from evalscope.config import TaskConfig
-from evalscope.constants import AnswerKeys, DumpMode, EvalStage, EvalType, ReviewKeys
+from evalscope.constants import AnswerKeys, DumpMode, EvalStage, EvalType, JudgeStrategy, ReviewKeys
 from evalscope.models import BaseModelAdapter
 from evalscope.report import Report, gen_table
 from evalscope.utils import dict_torch_dtype_to_str, gen_hash
@@ -58,9 +58,17 @@ class Evaluator(object):
         self.task_cfg = task_cfg
         # Deal with the output paths
         self.outputs_structure = outputs
-
         self.kwargs = kwargs
 
+        self._init_judge()
+
+    def _init_judge(self):
+        if self.task_cfg.judge_strategy == JudgeStrategy.RULE:
+            self.judge = None
+        else:
+            from evalscope.metrics.llm_judge import LLMJudge
+            self.judge = LLMJudge(**self.task_cfg.judge_model_args)
+
     def load_dataset(self):
         dataset = self.data_adapter.load(
             work_dir=os.path.expanduser(self.task_cfg.dataset_dir), datasets_hub=self.dataset_hub, **self.kwargs)
@@ -73,7 +81,7 @@ class Evaluator(object):
         for subset_name, prompts_list in prompts.items():
             limit = self.task_cfg.limit or len(prompts_list)
             for index, prompt in enumerate(prompts_list[:limit]):
-                prompt['index'] = index
+                prompt[AnswerKeys.INDEX] = index
                 limited_prompts[subset_name].append(prompt)
 
         return limited_prompts
@@ -89,7 +97,8 @@ class Evaluator(object):
         answer_d[AnswerKeys.ANSWER_ID] = answer_id
         answer_d[AnswerKeys.SUBSET_NAME] = subset_name
         answer_d[AnswerKeys.RAW_INPUT] = input_d[AnswerKeys.RAW_INPUT]
-        answer_d[AnswerKeys.ORIGIN_PROMPT] = input_d
+        # answer_d[AnswerKeys.ORIGIN_PROMPT] = input_d
+        answer_d[AnswerKeys.INDEX] = input_d[AnswerKeys.INDEX]
         return answer_d
 
     def _get_answer(self, input_prompts, subset_name, infer_cfg) -> List[dict]:
@@ -109,7 +118,7 @@ class Evaluator(object):
             return answers_list, prompts_list
 
         def get_answered_indices(answers_list: List[Dict]) -> List[int]:
-            indices = [answer[AnswerKeys.ORIGIN_PROMPT].get('index') for answer in answers_list]
+            indices = [answer.get(AnswerKeys.INDEX) for answer in answers_list]
 
             if all(index is None for index in indices):
                 return list(range(len(answers_list)))
@@ -200,17 +209,40 @@ class Evaluator(object):
         for choice in choices:
             raw_input_d: dict = review_res[AnswerKeys.RAW_INPUT]
             answer_content = choice[ReviewKeys.MESSAGE][ReviewKeys.CONTENT]
-            answer_content = self.data_adapter.parse_pred_result(
-                result=answer_content, raw_input_d=raw_input_d, eval_type=self.eval_type)
             gold_content = self.data_adapter.get_gold_answer(raw_input_d)
 
-            review_result = self.data_adapter.match(gold_content, answer_content)
+            # Get review result based on judge strategy
+            use_llm = (
+                self.task_cfg.judge_strategy == JudgeStrategy.LLM
+                or (self.task_cfg.judge_strategy == JudgeStrategy.AUTO and self.data_adapter.llm_as_a_judge))
+
+            if use_llm:
+                # Use LLM as judge
+                assert self.judge is not None, f'Judge model is required for LLM judging {self.data_adapter.name}'
+                review_result = self.data_adapter.llm_match(
+                    gold_content, answer_content, self.judge, raw_input=raw_input_d)
+                pred = answer_content
+            else:
+                # Use rule-based judging
+                pred_content = self.data_adapter.parse_pred_result(
+                    result=answer_content, raw_input_d=raw_input_d, eval_type=self.eval_type)
+                review_result = self.data_adapter.match(gold_content, pred_content)
+
+                # For LLM_RECALL strategy, use LLM to re-judge if rule-based result is not good
+                if (self.task_cfg.judge_strategy == JudgeStrategy.LLM_RECALL
+                        and isinstance(review_result, (bool, int, float)) and not bool(review_result)):
+                    assert self.judge is not None, f'Judge model is required for LLM_RECALL strategy {self.data_adapter.name}'  # noqa: E501
+                    review_result = self.data_adapter.llm_match(
+                        gold_content, answer_content, self.judge, raw_input=raw_input_d)
+                    pred = answer_content
+                else:
+                    pred = pred_content
+
             choice[ReviewKeys.REVIEW] = {
-                ReviewKeys.GOLD: gold_content,
-                ReviewKeys.PRED: answer_content,
+                ReviewKeys.GOLD: gold_content if gold_content != raw_input_d else '*Same as Input*',
+                ReviewKeys.PRED: pred,
                 ReviewKeys.RESULT: review_result
             }
-
             rev_choices.append(choice)
 
         review_res[AnswerKeys.CHOICES] = rev_choices
@@ -252,16 +284,23 @@ class Evaluator(object):
             logger.warning(f'Ignore use_cache={self.use_cache}, updating the review file: {review_file_path} ...')
             os.remove(review_file_path)
 
-        for answer_d in tqdm(answers_list, total=len(answers_list), desc=f'Reviewing({subset_name}): '):
+        def process_single_review(answer_d):
             review_id, reviewer_spec = self._generate_review_id(answer_d)
             # Get review
             review_d = self._get_review(answer_d=answer_d, review_id=review_id, reviewer_spec=reviewer_spec)
-
             logger.debug(review_d)
-
-            reviews_list.append(review_d)
-            # Dump reviews
-            dump_jsonl_data(review_d, review_file_path, dump_mode=DumpMode.APPEND)
+            return review_d
+
+        with ThreadPoolExecutor(max_workers=self.task_cfg.judge_worker_num) as executor:
+            # Submit all tasks and get futures
+            futures = [executor.submit(process_single_review, answer_d) for answer_d in answers_list]
+
+            # Process completed futures with progress bar
+            for future in tqdm(as_completed(futures), total=len(futures), desc=f'Reviewing({subset_name}): '):
+                review_d = future.result()
+                reviews_list.append(review_d)
+                # Dump reviews
+                dump_jsonl_data(review_d, review_file_path, dump_mode=DumpMode.APPEND)
 
         return reviews_list
 

evalscope/metrics/llm_judge.py

@@ -0,0 +1,104 @@
+import os
+import re
+from typing import Any, Dict, List, Optional
+
+from evalscope.utils.logger import get_logger
+
+logger = get_logger()
+
+DEFAULT_PROMPT_TEMPLATE = """Your job is to look at a question, a gold target, and a predicted answer, and return a letter "A" or "B" to indicate whether the predicted answer is correct or incorrect.
+
+Question: {question}
+
+Reference Answer: {gold}
+
+Model Answer: {pred}
+
+Evaluate the model's answer based on correctness compared to the reference answer.
+Grade the predicted answer of this new question as one of:
+A: CORRECT
+B: INCORRECT
+
+Just return the letters "A" or "B", with no text around it.
+"""  # noqa: E501
+
+
+class LLMJudge:
+    """
+    A metric that uses LLM to judge the quality of model predictions by comparing them with reference answers.
+    """
+
+    def __init__(self,
+                 api_key: Optional[str] = None,
+                 api_url: Optional[str] = None,
+                 model_id: Optional[str] = None,
+                 system_prompt: Optional[str] = None,
+                 prompt_template: Optional[str] = None,
+                 generation_config: Optional[Dict[str, Any]] = None,
+                 **kwargs):
+        """
+        Initialize LLMJudge metric.
+
+        Args:
+            api_key (str, optional): API key for OpenAI or compatible service
+            api_base (str, optional): API base URL
+            model_id (str, optional): Model ID for LLM
+            system_prompt (str, optional): System prompt for the judge
+            prompt_template (str, optional): Prompt template for the judge
+            generation_config (dict, optional): Generation configuration for the judge
+        """
+        self.api_key = api_key or os.environ.get('OPENAI_API_KEY', 'EMPTY')
+        self.api_url = api_url or os.environ.get('OPENAI_API_BASE', 'https://api.openai.com/v1')
+        self.model_id = model_id or os.environ.get('LOCAL_LLM', 'gpt-3.5-turbo')
+        self.system_prompt = system_prompt or os.environ.get('JUDGE_SYSTEM_PROMPT', None)
+        self.prompt_template = prompt_template or os.environ.get('JUDGE_PROMPT_TEMPLATE', DEFAULT_PROMPT_TEMPLATE)
+        self.generation_config = generation_config
+
+        from evalscope.models.server_adapter import ServerModelAdapter
+
+        # Initialize ServerModelAdapter
+        self.server_adapter = ServerModelAdapter(api_url=self.api_url, model_id=self.model_id, api_key=self.api_key)
+
+    def __call__(self, prompt: str, system_prompt: Optional[str] = None) -> float:
+        """
+        Args:
+            prompt (str): The prompt to evaluate
+            system_prompt (str, optional): The system prompt to use for the evaluation
+        Returns:
+            float: The score of the evaluation
+        """
+        input_data = {'data': [prompt], 'system_prompt': system_prompt or self.system_prompt}
+
+        # Inference configuration
+        infer_cfg = {'temperature': 0.0, 'max_tokens': 1024}
+        if self.generation_config:
+            infer_cfg.update(self.generation_config)
+
+        try:
+            # Send request using ServerModelAdapter
+            response = self.server_adapter.process_single_input(input_data, infer_cfg)
+
+            # Extract content from response
+            llm_response = response.get('choices', [{}])[0].get('message', {}).get('content', '')
+            return llm_response
+        except Exception as e:
+            logger.error(f'Error during LLM evaluation: {e}')
+            return None
+
+    def build_prompt(self, pred: str, gold: str, question: Optional[str] = None):
+        if question is None:
+            question = 'Not provided'
+        return self.prompt_template.format(question=question, pred=pred, gold=gold)
+
+    def get_score(self, response: str) -> float:
+        if response is None:
+            return 0
+        match = re.search(r'(A|B)', response)
+        if match:
+            answer = match.group(0)
+            if answer == 'A':
+                return 1
+            elif answer == 'B':
+                return 0
+        else:
+            return 0

evalscope/models/custom_adapter.py

@@ -66,4 +66,4 @@ class CustomModelAdapter(BaseModelAdapter):
             else:
                 raise TypeError(f'Unsupported inputs type: {type(input_prompt)}')
 
-        return self.custom_model.predict(prompts=in_prompts, **kwargs)
+        return self.custom_model.predict(prompts=in_prompts, origin_inputs=inputs, **kwargs)

evalscope/perf/arguments.py

@@ -24,6 +24,7 @@ class Arguments:
     connect_timeout: int = 600  # Connection timeout in seconds
     read_timeout: int = 600  # Read timeout in seconds
     api_key: Optional[str] = None
+    no_test_connection: bool = False  # Test the connection before starting the benchmark
 
     # Performance and parallelism
     number: Optional[int] = None  # Number of requests to be made
@@ -40,8 +41,9 @@ class Arguments:
     outputs_dir: str = DEFAULT_WORK_DIR
 
     # Prompt settings
-    max_prompt_length: int = sys.maxsize  # Maximum length of the prompt
+    max_prompt_length: int = 131072  # Maximum length of the prompt
     min_prompt_length: int = 0  # Minimum length of the prompt
+    prefix_length: int = 0  # Length of the prefix, only for random dataset
     prompt: Optional[str] = None  # The prompt text
     query_template: Optional[str] = None  # Template for the query
 
@@ -65,44 +67,12 @@ class Arguments:
 
     @staticmethod
     def from_args(args):
-        return Arguments(
-            model=args.model,
-            attn_implementation=args.attn_implementation,
-            url=args.url,
-            port=args.port,
-            api_key=args.api_key,
-            connect_timeout=args.connect_timeout,
-            read_timeout=args.read_timeout,
-            number=args.number,
-            parallel=args.parallel,
-            rate=args.rate,
-            log_every_n_query=args.log_every_n_query,
-            headers=args.headers,
-            wandb_api_key=args.wandb_api_key,
-            name=args.name,
-            outputs_dir=args.outputs_dir,
-            debug=args.debug,
-            tokenizer_path=args.tokenizer_path,
-            api=args.api,
-            max_prompt_length=args.max_prompt_length,
-            min_prompt_length=args.min_prompt_length,
-            prompt=args.prompt,
-            query_template=args.query_template,
-            dataset=args.dataset,
-            dataset_path=args.dataset_path,
-            frequency_penalty=args.frequency_penalty,
-            logprobs=args.logprobs,
-            max_tokens=args.max_tokens,
-            min_tokens=args.min_tokens,
-            n_choices=args.n_choices,
-            seed=args.seed,
-            stop=args.stop,
-            stop_token_ids=args.stop_token_ids,
-            stream=args.stream,
-            temperature=args.temperature,
-            top_p=args.top_p,
-            top_k=args.top_k,
-        )
+        # Convert Namespace to a dictionary and filter out None values
+        args_dict = {k: v for k, v in vars(args).items() if v is not None}
+
+        if 'func' in args_dict:
+            del args_dict['func']  # Note: compat CLI arguments
+        return Arguments(**args_dict)
 
     def __post_init__(self):
         self.headers = self.headers or {}  # Default to empty dictionary
@@ -153,6 +123,7 @@ def add_argument(parser: argparse.ArgumentParser):
     parser.add_argument('--api-key', type=str, required=False, default=None, help='The API key for authentication')
     parser.add_argument('--connect-timeout', type=int, default=600, help='The network connection timeout')
     parser.add_argument('--read-timeout', type=int, default=600, help='The network read timeout')
+    parser.add_argument('--no-test-connection', action='store_false', default=False, help='Do not test the connection before starting the benchmark')  # noqa: E501
 
     # Performance and parallelism
     parser.add_argument('-n', '--number', type=int, default=None, help='How many requests to be made')
@@ -168,6 +139,7 @@ def add_argument(parser: argparse.ArgumentParser):
     # Prompt settings
     parser.add_argument('--max-prompt-length', type=int, default=sys.maxsize, help='Maximum input prompt length')
     parser.add_argument('--min-prompt-length', type=int, default=0, help='Minimum input prompt length')
+    parser.add_argument('--prefix-length', type=int, default=0, help='The prefix length')
     parser.add_argument('--prompt', type=str, required=False, default=None, help='Specified the request prompt')
     parser.add_argument('--query-template', type=str, default=None, help='Specify the query template')
 
@@ -193,7 +165,6 @@ def add_argument(parser: argparse.ArgumentParser):
     parser.add_argument('--temperature', type=float, help='The sample temperature', default=None)
     parser.add_argument('--top-p', type=float, help='Sampling top p', default=None)
     parser.add_argument('--top-k', type=int, help='Sampling top k', default=None)
-
     # yapf: enable