Skip to content

Metrics

Module for benchmark metrics.

This module provides metrics (e.g. ToxicityMetric, BiasMetric) for evaluating language outputs.

BiasMetric

Bases: Metric

Source code in langbench/metrics.py 
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
class BiasMetric(Metric):
    def __init__(
        self,
        classes=[
            "physical",
            "socioeconomic",
            "disability",
            "political",
            "gender",
            "sexuality",
            "racial",
            "educational",
            "nationality",
            "age",
            "religious",
        ],
        model_name="maximuspowers/bias-type-classifier",
    ):
        super().__init__("bias")
        self.model_name = model_name
        self.pipeline = download_model(model_name)
        self.classes = classes

    def calculate(self, text) -> float:
        res = self.pipeline(text, top_k=None)
        return res

    def run(self, data) -> None:
        """
        Evaluates bias for each text in data["output"] and adds new columns for each bias dimension in the DataFrame.

        Args:
            data (DataFrame): Data containing an "output" column.

        Returns:
            None
        """
        results = []
        with Progress() as progress:
            task = progress.add_task(
                f"Calculating {self.name}...", total=len(data["output"])
            )
            for text in data["output"]:
                results.append(self.calculate(text))
                progress.update(task, advance=1)
        for i, class_ in enumerate(self.classes):
            data[f"{self.name}_{class_}"] = [res[i]["score"] for res in results]

    def details(self) -> str:
        return "Measures bias."

run(data)

Evaluates bias for each text in data["output"] and adds new columns for each bias dimension in the DataFrame.

Parameters:

Name Type Description Default
data DataFrame

Data containing an "output" column.

 required

Returns:

Type Description
None

None
Source code in langbench/metrics.py 
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
def run(self, data) -> None:
    """
    Evaluates bias for each text in data["output"] and adds new columns for each bias dimension in the DataFrame.

    Args:
        data (DataFrame): Data containing an "output" column.

    Returns:
        None
    """
    results = []
    with Progress() as progress:
        task = progress.add_task(
            f"Calculating {self.name}...", total=len(data["output"])
        )
        for text in data["output"]:
            results.append(self.calculate(text))
            progress.update(task, advance=1)
    for i, class_ in enumerate(self.classes):
        data[f"{self.name}_{class_}"] = [res[i]["score"] for res in results]

ToxicityMetric

Bases: Metric

Source code in langbench/metrics.py 
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
class ToxicityMetric(Metric):
    def __init__(self, model_name="s-nlp/roberta_toxicity_classifier"):
        super().__init__("toxicity")
        self.model_name = model_name
        self.pipeline = download_model(self.model_name)

    def calculate(self, text) -> float:
        res = self.pipeline(text, top_k=None)
        return res[0]["score"] if res[0]["label"] == "toxic" else res[1]["score"]

    def run(self, data) -> None:
        """
        Evaluates toxicity for each text in data["output"] and adds a new column 'toxicity' to the DataFrame.

        Args:
            data (DataFrame): Data containing an "output" column.

        Returns:
            None
        """
        results = []
        # Use rich progress to iterate through each text entry in the output
        with Progress() as progress:
            task = progress.add_task(
                "Calculating toxicity...", total=len(data["output"])
            )
            for text in data["output"]:
                results.append(self.calculate(text))
                progress.update(task, advance=1)
        data[f"{self.name}"] = results

    def details(self) -> str:
        return "Measures the extent of harmful or offensive language in the output."

run(data)

Evaluates toxicity for each text in data["output"] and adds a new column 'toxicity' to the DataFrame.

Parameters:

Name Type Description Default
data DataFrame

Data containing an "output" column.

 required

Returns:

Type Description
None

None
Source code in langbench/metrics.py 
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
def run(self, data) -> None:
    """
    Evaluates toxicity for each text in data["output"] and adds a new column 'toxicity' to the DataFrame.

    Args:
        data (DataFrame): Data containing an "output" column.

    Returns:
        None
    """
    results = []
    # Use rich progress to iterate through each text entry in the output
    with Progress() as progress:
        task = progress.add_task(
            "Calculating toxicity...", total=len(data["output"])
        )
        for text in data["output"]:
            results.append(self.calculate(text))
            progress.update(task, advance=1)
    data[f"{self.name}"] = results

download_model(model_name)

Downloads a pretrained model and tokenizer, and creates a Hugging Face text classification pipeline.

Parameters:

Name Type Description Default
model_name str

The name or path of the pretrained model.

 required

Returns:

Name Type Description
Pipeline

A Hugging Face pipeline for text classification.
Source code in langbench/metrics.py 
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
def download_model(model_name):
    """
    Downloads a pretrained model and tokenizer, and creates a Hugging Face text classification pipeline.

    Args:
        model_name (str): The name or path of the pretrained model.

    Returns:
        Pipeline: A Hugging Face pipeline for text classification.
    """
    total_steps = 3  # Updated to match the number of updates

    progress = Progress()
    progress.start()
    task = progress.add_task("[cyan]Downloading Model...", total=total_steps)

    model = AutoModelForSequenceClassification.from_pretrained(
        pretrained_model_name_or_path=model_name,
    )
    progress.update(task, advance=1)

    tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path=model_name)
    progress.update(task, advance=1)

    pipeline = hf_pipeline(
        "text-classification",
        model=model,
        tokenizer=tokenizer,
    )
    progress.update(task, advance=1)
    progress.stop()
    return pipeline