📄 Long Bench V2¶

LongBench v2 is designed to assess the ability of LLMs to handle long-context problems requiring deep understanding and reasoning across real-world multitasks. LongBench v2 has the following features: (1) Length: Context length ranging from 8k to 2M words, with the majority under 128k. (2) Difficulty: Challenging enough that even human experts, using search tools within the document, cannot answer correctly in a short time. (3) Coverage: Cover various realistic scenarios. (4) Reliability: All in a multiple-choice question format for reliable evaluation.

To elaborate, LongBench v2 consists of 503 challenging multiple-choice questions, with contexts ranging from 8k to 2M words, across six major task categories: single-document QA, multi-document QA, long in-context learning, long-dialogue history understanding, code repo understanding, and long structured data understanding. To ensure the breadth and the practicality, we collect data from nearly 100 highly educated individuals with diverse professional backgrounds. We employ both automated and manual review processes to maintain high quality and difficulty, resulting in human experts achieving only 53.7% accuracy under a 15-minute time constraint. Our evaluation reveals that the best-performing model, when directly answers the questions, achieves only 50.1% accuracy. In contrast, the o1-preview model, which includes longer reasoning, achieves 57.7%, surpassing the human baseline by 4%. These results highlight the importance of enhanced reasoning ability and scaling inference-time compute to tackle the long-context challenges in LongBench v2.

Tags: annotations_creators:expert-generated, arxiv:2412.15204, flags:['NLU', 'natural language understanding'], language:en, language_creators:other, license:other, multilinguality:monolingual, region:us, size_categories:n<1K, source_datasets:extended|other, task_categories:['text-classification', 'token-classification', 'question-answering'], task_ids:['natural-language-inference', 'word-sense-disambiguation', 'coreference-resolution', 'extractive-qa'], category:dataset

cards.long_bench_v2

TaskCard(
    loader=LoadHF(
        path="THUDM/LongBench-v2",
        data_classification_policy=[
            "public",
        ],
    ),
    preprocess_steps=[
        RenameSplits(
            mapper={
                "train": "test",
            },
        ),
        ListFieldValues(
            fields=[
                "choice_A",
                "choice_B",
                "choice_C",
                "choice_D",
            ],
            to_field="choices",
        ),
        Copy(
            field="domain",
            to_field="context_type",
        ),
        MapInstanceValues(
            mappers={
                "answer": {
                    "A": 0,
                    "B": 1,
                    "C": 2,
                    "D": 3,
                },
                "context_type": {
                    "Long In-context Learning": "examples",
                    "Single-Document QA": "document",
                    "Long Structured Data Understanding": "data",
                    "Multi-Document QA": "documents",
                    "Code Repository Understanding": "code",
                    "Long-dialogue History Understanding": "dialog",
                },
            },
        ),
    ],
    task="tasks.qa.multiple_choice.with_context",
    templates="templates.qa.multiple_choice.with_context.all",
)

[source]

from unitxt.loaders import LoadHF
from unitxt.operators import Copy, ListFieldValues, MapInstanceValues
from unitxt.splitters import RenameSplits

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 MapInstanceValues¶

A class used to map instance values into other values.

This class is a type of InstanceOperator, it maps values of instances in a stream using predefined mappers.

Args:

mappers (Dict[str, Dict[str, Any]]):
The mappers to use for mapping instance values. Keys are the names of the fields to undergo mapping, and values are dictionaries that define the mapping from old values to new values. Note that mapped values are defined by their string representation, so mapped values are converted to strings before being looked up in the mappers.

strict (bool):
If True, the mapping is applied strictly. That means if a value does not exist in the mapper, it will raise a KeyError. If False, values that are not present in the mapper are kept as they are.

process_every_value (bool):
If True, all fields to be mapped should be lists, and the mapping is to be applied to their individual elements. If False, mapping is only applied to a field containing a single value.

Examples:
MapInstanceValues(mappers={"a": {"1": "hi", "2": "bye"}}) replaces "1" with "hi" and "2" with "bye" in field "a" in all instances of all streams: instance {"a": 1, "b": 2} becomes {"a": "hi", "b": 2}. Note that the value of "b" remained intact, since field-name "b" does not participate in the mappers, and that 1 was casted to "1" before looked up in the mapper of "a".

MapInstanceValues(mappers={"a": {"1": "hi", "2": "bye"}}, process_every_value=True): Assuming field "a" is a list of values, potentially including "1"-s and "2"-s, this replaces each such "1" with "hi" and "2" – with "bye" in all instances of all streams: instance {"a": ["1", "2"], "b": 2} becomes {"a": ["hi", "bye"], "b": 2}.

MapInstanceValues(mappers={"a": {"1": "hi", "2": "bye"}}, strict=True): To ensure that all values of field "a" are mapped in every instance, use strict=True. Input instance {"a":"3", "b": 2} will raise an exception per the above call, because "3" is not a key in the mapper of "a".

MapInstanceValues(mappers={"a": {str([1,2,3,4]): "All", str([]): "None"}}, strict=True) replaces a list [1,2,3,4] with the string "All" and an empty list by string "None".

Explanation about Copy¶

Copies values from specified fields to specified fields.

Args (of parent class):
field_to_field (Union[List[List], Dict[str, str]]): A list of lists, where each sublist contains the source field and the destination field, or a dictionary mapping source fields to destination fields.

Examples:
An input instance {“a”: 2, “b”: 3}, when processed by Copy(field_to_field={"a": "b"}) would yield {“a”: 2, “b”: 2}, and when processed by Copy(field_to_field={"a": "c"}) would yield {“a”: 2, “b”: 3, “c”: 2}

with field names containing / , we can also copy inside the field: Copy(field="a/0",to_field="a") would process instance {“a”: [1, 3]} into {“a”: 1}

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 ListFieldValues¶

Concatenates values of multiple fields into a list, and assigns it to a new field.

References: templates.qa.multiple_choice.with_context.all, tasks.qa.multiple_choice.with_context