multilingual.py

import os
import json
import jinja2

import datasets

logger = datasets.logging.get_logger(__name__)

_LANG = ["ar", "en", "en-ar"]
_COLLECTION = ["ncwm", "ncwm-1000", "ncwm-5000", "ncwm-10000", "adgen", "dialog", "arce", "alpaca"]


class JinJa2Formatter:
    def __init__(self, instruction: str, input: str, output=""):
        self.instruction = jinja2.Template(instruction)
        self.input = jinja2.Template(input)
        self.output = jinja2.Template(output)

    def __call__(self, example):
        try:
            return {
                "instruction": self.instruction.render(**example),
                "input": self.input.render(**example),
                "output": self.output.render(**example),
            }
        except Exception as e:
            raise ValueError(f"Error while formatting example: {example}") from e


_FORMATTER = {
    "adgen": JinJa2Formatter(
        instruction="Generate advertisement for product according to its description, using the language provided in the contents.",
        input="Product:{{product}}\nDescription:{{description}}",
        output="{{ad}}",
    ),
    "dialog": JinJa2Formatter(
        instruction="Summarize the dialogue with respect to the provided topic. Use <end> to end your response",
        input="Dialogue:{{dialogue}}\nTopic:{{topic}}",
        output="{{summary}}",
    ),
    "arce": JinJa2Formatter(
        instruction="Question:{{question}}\nChoices:{{choices.text}}",
        input="",
        output="{{choices.text[choices.label.index(answerKey)]}}",
    ),
    "ncwm": JinJa2Formatter(
        instruction="{{instruction}}",
        input="{{input}}",
        output="{{output}}",
    ),
    "alpaca": JinJa2Formatter(
        instruction="{{instruction}}",
        input="{{input}}",
        output="{{output}}",
    ),
}

_FORMATTER["ncwm-1000"] = _FORMATTER["ncwm"]
_FORMATTER["ncwm-5000"] = _FORMATTER["ncwm"]
_FORMATTER["ncwm-10000"] = _FORMATTER["ncwm"]


class MultilingualConfig(datasets.BuilderConfig):
    """BuilderConfig for Alpaca"""

    def __init__(self, lang: str, collection: str, **kwargs):
        """
        Args:
            lang: string, language for the input text
            collection: string, collection name
            **kwargs: keyword arguments forwarded to super.
        """
        super(MultilingualConfig, self).__init__(**kwargs)
        self.lang = lang
        self.collection = collection


def _get_config(collection, lang):
    return MultilingualConfig(lang=lang, collection=collection, name=f"{collection}_{lang}")


class Multilingual(datasets.GeneratorBasedBuilder):
    VERSION = datasets.Version("1.0.0")
    BUILDER_CONFIGS = [
        _get_config("adgen", "ar"),
        _get_config("adgen", "en"),
        _get_config("dialog", "ar"),
        _get_config("dialog", "en"),
        _get_config("arce", "en"),
        _get_config("arce", "ar"),
        _get_config("ncwm", "en-ar"),
        _get_config("ncwm-1000", "en-ar"),
        _get_config("ncwm-5000", "en-ar"),
        _get_config("ncwm-10000", "en-ar"),
        _get_config("alpaca", "en"),
    ]
    BUILDER_CONFIG_CLASS = MultilingualConfig

    def _info(self):
        return datasets.DatasetInfo(
            features=datasets.Features(
                {
                    "id": datasets.Value("string"),
                    "instruction": datasets.Value("string"),
                    "input": datasets.Value("string"),
                    "output": datasets.Value("string"),
                }
            ),
        )

    def _split_generators(self, dl_manager):
        splits_generators = []
        for name in [
            datasets.Split.TRAIN,
            datasets.Split.TEST,
            datasets.Split.VALIDATION,
        ]:
            filepath = os.path.join(
                self.base_path,
                f"{self.config.collection}_{self.config.lang}_{name}.jsonl",
            )
            if os.path.exists(filepath):
                splits_generators.append(datasets.SplitGenerator(name=name, gen_kwargs={"filepath": filepath}))
        if not splits_generators:
            raise ValueError("no splits found")
        return splits_generators

    def _generate_examples(self, filepath):
        """This function returns the examples in the raw (text) form."""
        logger.info("[multilingual] generating examples from = %s", filepath)

        formatter = None
        if f"{self.config.collection}_{self.config.lang}" in _FORMATTER:
            formatter = _FORMATTER[f"{self.config.collection}_{self.config.lang}"]
        elif f"{self.config.collection}" in _FORMATTER:
            formatter = _FORMATTER[f"{self.config.collection}"]
        else:
            raise ValueError(
                f"Formatter for the collection `{self.config.collection}` and language `{self.config.lang}` not found."
            )

        with open(filepath, encoding="utf-8") as f:
            samples = [json.loads(x) for x in f.readlines()]
            id_ = 0
            for sample in samples:
                yield id_, formatter(sample) | {"id": str(id_)}
                id_ += 1