📄 REALMMRAG: TechReport¶

We introduced REAL-MM-RAG-Bench, a real-world multi-modal retrieval benchmark designed to evaluate retrieval models in reliable, challenging, and realistic settings. The benchmark was constructed using an automated pipeline, where queries were generated by a vision-language model (VLM), filtered by a large language model (LLM), and rephrased by an LLM to ensure high-quality retrieval evaluation. To simulate real-world retrieval challenges, we introduce multi-level query rephrasing, modifying queries at three distinct levels—from minor wording adjustments to significant structural changes—ensuring models are tested on their true semantic understanding rather than simple keyword matching.

Tags: license:cdla-permissive-2.0, url:https://huggingface.co/datasets//ibm-research/REAL-MM-RAG_TechReport, category:dataset

cards.rag.benchmark.real_mm_rag_tech_report.en

TaskCard(
    loader=LoadHF(
        path="ibm-research/REAL-MM-RAG_TechReport",
        name="default",
        split="test",
        data_classification_policy=[
            "public",
        ],
    ),
    preprocess_steps=[
        FilterByCondition(
            values={
                "query": None,
            },
            condition="ne",
        ),
        HashImage(
            field="image",
            to_field="reference_context_ids",
        ),
        Copy(
            field="query",
            to_field="question",
        ),
        AddIncrementalId(
            to_field="question_id",
        ),
        Cast(
            field="question_id",
            to="str",
        ),
        SplitRandomMix(
            mix={
                "test": "test[30%]",
                "train": "test[70%]",
            },
        ),
        Wrap(
            field="answer",
            inside="list",
            to_field="reference_answers",
        ),
        Wrap(
            field="reference_context_ids",
            inside="list",
            to_field="reference_context_ids",
        ),
    ],
    task="tasks.rag.end_to_end",
    templates={
        "default": "templates.rag.end_to_end.json_predictions",
    },
    __title__="REALMMRAG: TechReport",
)

[source]

from unitxt.collections_operators import Wrap
from unitxt.image_operators import HashImage
from unitxt.loaders import LoadHF
from unitxt.operators import AddIncrementalId, Cast, Copy, FilterByCondition
from unitxt.splitters import SplitRandomMix

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 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 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 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 Cast¶

Casts specified fields to specified types.

Args:
default (object): A dictionary mapping field names to default values for cases of casting failure. process_every_value (bool): If true, all fields involved must contain lists, and each value in the list is then casted. Defaults to False.

Explanation about SplitRandomMix¶

Splits a multistream into new streams (splits), whose names, source input stream, and amount of instances, are specified by arg ‘mix’.

The keys of arg ‘mix’, are the names of the new streams, the values are of the form: ‘name-of-source-stream[percentage-of-source-stream]’ Each input instance, of any input stream, is selected exactly once for inclusion in any of the output streams.

Examples: When processing a multistream made of two streams whose names are ‘train’ and ‘test’, by SplitRandomMix(mix = { “train”: “train[99%]”, “validation”: “train[1%]”, “test”: “test” }) the output is a multistream, whose three streams are named ‘train’, ‘validation’, and ‘test’. Output stream ‘train’ is made of randomly selected 99% of the instances of input stream ‘train’, output stream ‘validation’ is made of the remaining 1% instances of input ‘train’, and output stream ‘test’ is made of the whole of input stream ‘test’.

When processing the above input multistream by SplitRandomMix(mix = { “train”: “train[50%]+test[0.1]”, “validation”: “train[50%]+test[0.2]”, “test”: “test[0.7]” }) the output is a multistream, whose three streams are named ‘train’, ‘validation’, and ‘test’. Output stream ‘train’ is made of randomly selected 50% of the instances of input stream ‘train’ + randomly selected 0.1 (i.e., 10%) of the instances of input stream ‘test’. Output stream ‘validation’ is made of the remaining 50% instances of input ‘train’+ randomly selected 0.2 (i.e., 20%) of the original instances of input ‘test’, that were not selected for output ‘train’, and output stream ‘test’ is made of the remaining instances of input ‘test’.

References: templates.rag.end_to_end.json_predictions, tasks.rag.end_to_end