dataset_dict.py

import os
import json
import datasets

class MixBench(datasets.GeneratorBasedBuilder):
    def _info(self):
        return datasets.DatasetInfo(
            features=datasets.Features({
                "query_id": datasets.Value("string"),
                "query_text": datasets.Value("string"),
                "query_image": datasets.Value("string"),
                "corpus_id": datasets.Value("string"),
                "corpus_text": datasets.Value("string"),
                "corpus_image": datasets.Value("string"),
                "score": datasets.Value("int32")
            })
        )

    def _split_generators(self, dl_manager):
        data_dir = dl_manager.download_and_extract(self.config.data_dir)
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"data_dir": data_dir}
            )
        ]

    def _generate_examples(self, data_dir):
        with open(os.path.join(data_dir, "queries.jsonl")) as fq:
            queries = {q["query_id"]: q for q in map(json.loads, fq)}
        with open(os.path.join(data_dir, "corpus.jsonl")) as fc:
            corpus = {c["corpus_id"]: c for c in map(json.loads, fc)}
        with open(os.path.join(data_dir, "qrels.tsv")) as fr:
            for idx, line in enumerate(fr):
                qid, cid, score = line.strip().split("\t")
                yield idx, {
                    "query_id": qid,
                    "query_text": queries[qid].get("text", ""),
                    "query_image": queries[qid].get("image", ""),
                    "corpus_id": cid,
                    "corpus_text": corpus[cid].get("text", ""),
                    "corpus_image": corpus[cid].get("image", ""),
                    "score": int(score)
                }