BenchDatasetEvaluator Operator Reference

BenchDatasetEvaluator compares predicted answers against ground truth using two modes: match (math verification) or semantic (LLM-based).

1. Import

from dataflow.operators.core_text import BenchDatasetEvaluator

2. Match Mode

Constructor

BenchDatasetEvaluator(
    eval_result_path=None,
    compare_method="match",
)

run() Signature

op.run(
    storage=self.storage.step(),
    input_test_answer_key="generated_cot",
    input_gt_answer_key="golden_answer",
)
# returns: [input_test_answer_key, input_gt_answer_key, 'answer_match_result']

Runtime Logic

1. Read DataFrame from storage.
2. Create answer_match_result column initialized to False.
3. For each row, extract answer using AnswerExtractor and compare with ground truth using math_verify_compare().
4. Write results to answer_match_result column.
5. Save statistics to eval_result_path.
6. Return column list.

Usage Example

from dataflow.operators.core_text import BenchDatasetEvaluator
from dataflow.utils.storage import FileStorage

class MyPipeline:
    def __init__(self):
        self.storage = FileStorage(
            first_entry_file_name="./data/bench.jsonl",
            cache_path="./cache",
            file_name_prefix="step",
            cache_type="jsonl"
        )

        self.evaluator = BenchDatasetEvaluator(
            compare_method="match",
            eval_result_path="./results/match_eval.json"
        )

    def forward(self):
        self.evaluator.run(
            storage=self.storage.step(),
            input_test_answer_key="predicted_answer",
            input_gt_answer_key="ground_truth"
        )

if __name__ == "__main__":
    pipeline = MyPipeline()
    pipeline.forward()

3. Semantic Mode

Constructor

BenchDatasetEvaluator(
    eval_result_path=None,
    compare_method="semantic",
    system_prompt="You are a helpful assistant specialized in evaluating answer correctness.",
    llm_serving=llm_serving,
    prompt_template=None,
)

run() Signature

op.run(
    storage=self.storage.step(),
    input_test_answer_key="generated_cot",
    input_gt_answer_key="golden_answer",
)
# returns: [input_test_answer_key, input_gt_answer_key, 'answer_match_result']

Runtime Logic

1. Read DataFrame from storage.
2. Create answer_match_result column initialized to False.
3. Skip rows where ground truth is empty/NaN.
4. Build prompts using only answer and reference_answer fields.
5. Call LLM to judge correctness.
6. Parse LLM response for "judgement_result": true/false.
7. Write results to answer_match_result column.
8. Save statistics to eval_result_path.
9. Return column list.

Usage Example

from dataflow.operators.core_text import BenchDatasetEvaluator
from dataflow.serving import APILLMServing_request
from dataflow.utils.storage import FileStorage

class MyPipeline:
    def __init__(self):
        self.storage = FileStorage(
            first_entry_file_name="./data/bench.jsonl",
            cache_path="./cache",
            file_name_prefix="step",
            cache_type="jsonl"
        )

        self.llm_serving = APILLMServing_request(
            api_url="https://api.openai.com/v1/chat/completions",
            key_name_of_api_key="DF_API_KEY",
            model_name="gpt-4o",
            max_workers=10
        )

        self.evaluator = BenchDatasetEvaluator(
            compare_method="semantic",
            llm_serving=self.llm_serving,
            prompt_template=None,
            eval_result_path="./results/semantic_eval.json"
        )

    def forward(self):
        self.evaluator.run(
            storage=self.storage.step(),
            input_test_answer_key="predicted_answer",
            input_gt_answer_key="ground_truth"
        )

if __name__ == "__main__":
    pipeline = MyPipeline()
    pipeline.forward()

4. AnswerJudgePrompt

AnswerJudgePrompt is the default prompt template class for semantic mode.

Important Notes

1. Default value issue: Constructor default is prompt_template=AnswerJudgePrompt (class, not instance).
2. Recommended usage: Pass prompt_template=None to use built-in fallback.
3. Custom template: Pass an instance if you need custom prompts.

Prompt Structure

The prompt template builds a JSON-formatted request:

• answer: The predicted answer to evaluate
• reference_answer: The ground truth answer

LLM response must contain:

{
  "judgement_result": true  // or false
}

Custom Template Example

from dataflow.prompts.core_text import AnswerJudgePrompt

custom_prompt = AnswerJudgePrompt()
# Modify if needed

evaluator = BenchDatasetEvaluator(
    compare_method="semantic",
    llm_serving=llm_serving,
    prompt_template=custom_prompt
)