📄 Airbench2024¶

AIRBench 2024 is a AI safety benchmark that aligns with emerging government regulations and company policies. It consists of diverse, malicious prompts spanning categories of the regulation-based safety categories in the AIR 2024 safety taxonomy.

Tags: languages:['english'], urls:{'arxiv': 'https://arxiv.org/abs/2407.17436'}, category:dataset

cards.safety.airbench2024

TaskCard(
    loader=MultipleSourceLoader(
        sources=[
            LoadHF(
                path="stanford-crfm/air-bench-2024",
                name="default",
            ),
            LoadHF(
                path="stanford-crfm/air-bench-2024",
                data_files={
                    "judge_prompts": "judge_prompt_final.csv",
                },
            ),
        ],
        data_classification_policy=[
            "public",
        ],
    ),
    preprocess_steps=[
        SelectFields(
            fields=[
                "cate-idx",
                "judge_prompt",
            ],
            apply_to_streams=[
                "judge_prompts",
            ],
        ),
        JoinStreams(
            left_stream="test",
            right_stream="judge_prompts",
            how="inner",
            on=[
                "cate-idx",
            ],
            new_stream_name="test",
        ),
        DeleteSplits(
            splits=[
                "judge_prompts",
            ],
        ),
    ],
    task=Task(
        input_fields={
            "cate-idx": "str",
            "l2-name": "str",
            "l3-name": "str",
            "l4-name": "str",
            "prompt": "str",
            "judge_prompt": "str",
        },
        reference_fields={},
        prediction_type="str",
        metrics=[
            "metrics.granite_guardian.assistant_risk.harm[prediction_type=str,user_message_field=prompt,assistant_message_field=prediction]",
        ],
    ),
    templates=TemplatesDict(
        items={
            "default": InputOutputTemplate(
                input_format="{prompt}
",
                output_format="",
            ),
        },
    ),
)

[source]

from unitxt.loaders import LoadHF, MultipleSourceLoader
from unitxt.operators import SelectFields
from unitxt.stream_operators import DeleteSplits, JoinStreams
from unitxt.task import Task
from unitxt.templates import InputOutputTemplate, TemplatesDict

Explanation about TaskCard¶

TaskCard delineates the phases in transforming the source dataset into model input, and specifies the metrics for evaluation of model output.

Args:

loader:
specifies the source address and the loading operator that can access that source and transform it into a unitxt multistream.

preprocess_steps:
list of unitxt operators to process the data source into model input.

task:
specifies the fields (of the already (pre)processed instance) making the inputs, the fields making the outputs, and the metrics to be used for evaluating the model output.

templates:
format strings to be applied on the input fields (specified by the task) and the output fields. The template also carries the instructions and the list of postprocessing steps, to be applied to the model output.

Explanation about JoinStreams¶

Join multiple streams into a single stream.

Args:
left_stream (str): The stream that will be considered the “left” in the join operations. right_stream (str): The stream that will be considered the “right” in the join operations. how (Literal[“left”, “right”, “inner”, “outer”, “cross”]): The type of join to be performed. on (Optional[List[str]]): Column names to join on. These must be found in both streams. left_on (Optional[List[str]]): Column names to join on in the left stream. right_on (Optional[List[str]]): Column names to join on in the right streasm. new_stream_name (str): The name of the new stream resulting from the merge.

Examples:
JoinStreams(left_stream = “questions”, right_stream = “answers”, how=”inner”, on=”question_id”, new_stream_name=”question_with_answers” ) Join the ‘question’ and ‘answer’ stream based on the ‘question_id’ field using inner join, resulting with a new stream named “question_with_answers”. JoinStreams(left_stream = “questions”, right_stream = “answers”, how=”inner”, on_left=”question_id”, on_right=”question” new_stream_name=”question_with_answers” ) Join the ‘question’ and ‘answer’ stream based on the ‘question_id’ field in the left stream and the ‘question’ field in the right stream, using inner join, resulting with a new stream named “question_with_answers”. This is suitable when the fields have different labels across the streams.

Explanation about Task¶

Task packs the different instance fields into dictionaries by their roles in the task.

Args:

input_fields (Union[Dict[str, str], List[str]]):
Dictionary with string names of instance input fields and types of respective values. In case a list is passed, each type will be assumed to be Any.

reference_fields (Union[Dict[str, str], List[str]]):
Dictionary with string names of instance output fields and types of respective values. In case a list is passed, each type will be assumed to be Any.

metrics (List[str]):
List of names of metrics to be used in the task.

prediction_type (Optional[str]):
Need to be consistent with all used metrics. Defaults to None, which means that it will be set to Any.

defaults (Optional[Dict[str, Any]]):
An optional dictionary with default values for chosen input/output keys. Needs to be consistent with names and types provided in ‘input_fields’ and/or ‘output_fields’ arguments. Will not overwrite values if already provided in a given instance.

The output instance contains three fields:

“input_fields” whose value is a sub-dictionary of the input instance, consisting of all the fields listed in Arg ‘input_fields’.

“reference_fields” – for the fields listed in Arg “reference_fields”.

“metrics” – to contain the value of Arg ‘metrics’

Explanation about LoadHF¶

Loads datasets from the HuggingFace Hub.

It supports loading with or without streaming, and it can filter datasets upon loading.
Args:

path:
The path or identifier of the dataset on the HuggingFace Hub.

name:
An optional dataset name.

data_dir:
Optional directory to store downloaded data.

split:
Optional specification of which split to load.

data_files:
Optional specification of particular data files to load. When you provide a list of data_files to Hugging Face’s load_dataset function without explicitly specifying the split argument, these files are automatically placed into the train split.

revision:
Optional. The revision of the dataset. Often the commit id. Use in case you want to set the dataset version.

streaming (bool):
indicating if streaming should be used.

filtering_lambda (str, optional):
A lambda function for filtering the data after loading.

num_proc (int, optional):
Specifies the number of processes to use for parallel dataset loading.

Example:
Loading glue’s mrpc dataset
load_hf = LoadHF(path='glue', name='mrpc')

Explanation about SelectFields¶

Keep only specified fields from each instance in a stream.

Args:
fields (List[str]): The fields to keep from each instance.

Explanation about MultipleSourceLoader¶

Allows loading data from multiple sources, potentially mixing different types of loaders.

Args:
sources: A list of loaders that will be combined to form a unified dataset.

Examples:
Loading the train split from a HuggingFace Hub and the test set from a local file:
MultipleSourceLoader(sources = [ LoadHF(path="public/data",split="train"), LoadCSV({"test": "mytest.csv"}) ])
Loading a test set combined from two files
MultipleSourceLoader(sources = [ LoadCSV({"test": "mytest1.csv"}, LoadCSV({"test": "mytest2.csv"}) ])

Explanation about InputOutputTemplate¶

Generate field ‘source’ from fields designated as input, and fields ‘target’ and ‘references’ from fields designated as output, of the processed instance.

Args specify the formatting strings with which to glue together the input and reference fields of the processed instance into one string (‘source’ and ‘target’), and into a list of strings (‘references’).

Explanation about DeleteSplits¶

Operator which delete splits in stream.

Attributes:
splits (List[str]): The splits to delete from the stream.

References: metrics.granite_guardian.assistant_risk.harm