📄 Multi Turn Gpt4 Judgement¶

cards.mt_bench.response_assessment.rating.multi_turn_gpt4_judgement

TaskCard(
    loader=LoadFromHFSpace(
        space_name="lmsys/mt-bench",
        revision="a4b674c",
        data_files={
            "questions": "data/mt_bench/question.jsonl",
            "model_answer": "data/mt_bench/model_answer/*.jsonl",
            "judgment": "data/mt_bench/model_judgment/gpt-4_single.jsonl",
        },
        data_classification_policy=[
            "public",
        ],
    ),
    preprocess_steps=[
        "operators.mt_bench.rating_hf_space_processing_steps",
        FilterByCondition(
            values={
                "turn": 2,
            },
            condition="eq",
        ),
        Fillna(
            field="reference",
            value=None,
        ),
        FilterByCondition(
            values={
                "reference": None,
            },
            condition="eq",
        ),
        Rename(
            field_to_field={
                "score": "rating",
                "category": "group",
            },
        ),
        InterleaveListsToDialogOperator(
            user_turns_field="model_input",
            assistant_turns_field="model_output",
            to_field="dialog",
        ),
    ],
    task="tasks.response_assessment.rating.multi_turn",
    templates=[
        "templates.response_assessment.rating.mt_bench_multi_turn",
    ],
)

[source]

from unitxt.loaders import LoadFromHFSpace
from unitxt.operators import Fillna, FilterByCondition, InterleaveListsToDialogOperator, Rename

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 FilterByCondition¶

Filters a stream, yielding only instances in which the values in required fields follow the required condition operator.

Raises an error if a required field name is missing from the input instance.

Args:
values (Dict[str, Any]): Field names and respective Values that instances must match according the condition, to be included in the output.

condition: the name of the desired condition operator between the specified (sub) field’s value and the provided constant value. Supported conditions are (“gt”, “ge”, “lt”, “le”, “ne”, “eq”, “in”,”not in”)

error_on_filtered_all (bool, optional): If True, raises an error if all instances are filtered out. Defaults to True.

Examples:

FilterByCondition(values = {"a":4}, condition = "gt") will yield only instances where field "a" contains a value > 4

FilterByCondition(values = {"a":4}, condition = "le") will yield only instances where "a"<=4

FilterByCondition(values = {"a":[4,8]}, condition = "in") will yield only instances where "a" is 4 or 8

FilterByCondition(values = {"a":[4,8]}, condition = "not in") will yield only instances where "a" is different from 4 or 8

FilterByCondition(values = {"a/b":[4,8]}, condition = "not in") will yield only instances where "a" is a dict in which key "b" is mapped to a value that is neither 4 nor 8

FilterByCondition(values = {"a[2]":4}, condition = "le") will yield only instances where “a” is a list whose 3-rd element is <= 4

FilterByCondition(values = {"a":False}, condition = "exists") will yield only instances which do not contain a field named "a"

FilterByCondition(values = {"a/b":True}, condition = "exists") will yield only instances which contain a field named "a" whose value is a dict containing, in turn, a field named "b"

Explanation about LoadFromHFSpace¶

Used to load data from HuggingFace Spaces lazily.

Args:

space_name (str):
Name of the HuggingFace Space to be accessed.

data_files (str | Sequence[str] | Mapping[str, str | Sequence[str]]):
Relative paths to files within a given repository. If given as a mapping, paths should be values, while keys should represent the type of respective files (training, testing etc.).

path (str, optional):
Absolute path to a directory where data should be downloaded.

revision (str, optional):
ID of a Git branch or commit to be used. By default, it is set to None, thus data is downloaded from the main branch of the accessed repository.

use_token (bool, optional):
Whether a token is used for authentication when accessing the HuggingFace Space. If necessary, the token is read from the HuggingFace config folder.

token_env (str, optional):
Key of an env variable which value will be used for authentication when accessing the HuggingFace Space - if necessary.

Explanation about Rename¶

Renames fields.

Move value from one field to another, potentially, if field name contains a /, from one branch into another. Remove the from field, potentially part of it in case of / in from_field.

Examples:
Rename(field_to_field={“b”: “c”}) will change inputs [{“a”: 1, “b”: 2}, {“a”: 2, “b”: 3}] to [{“a”: 1, “c”: 2}, {“a”: 2, “c”: 3}]

Rename(field_to_field={“b”: “c/d”}) will change inputs [{“a”: 1, “b”: 2}, {“a”: 2, “b”: 3}] to [{“a”: 1, “c”: {“d”: 2}}, {“a”: 2, “c”: {“d”: 3}}]

Rename(field_to_field={“b”: “b/d”}) will change inputs [{“a”: 1, “b”: 2}, {“a”: 2, “b”: 3}] to [{“a”: 1, “b”: {“d”: 2}}, {“a”: 2, “b”: {“d”: 3}}]

Rename(field_to_field={“b/c/e”: “b/d”}) will change inputs [{“a”: 1, “b”: {“c”: {“e”: 2, “f”: 20}}}] to [{“a”: 1, “b”: {“c”: {“f”: 20}, “d”: 2}}]

Explanation about InterleaveListsToDialogOperator¶

Interleaves two lists, one of user dialog turns and one of assistant dialog turns, into a single list of tuples, alternating between “user” and “assistant”.

The list of tuples if of format (role, turn_content), where the role label is specified by the ‘user_role_label’ and ‘assistant_role_label’ fields (default to “user” and “assistant”).

The user turns and assistant turns field are specified in the arguments. The value of each of the ‘fields’ is assumed to be a list.

References: templates.response_assessment.rating.mt_bench_multi_turn, operators.mt_bench.rating_hf_space_processing_steps, tasks.response_assessment.rating.multi_turn