dataset.py

import json
import datasets

_DESCRIPTION = """
LEMAS Eval Dataset (Multilingual)

Each sample contains:
- audio: speech audio
- txt: normalized transcript
- align: word-level alignment (json string)
"""

class LEMASEval(datasets.GeneratorBasedBuilder):
    VERSION = datasets.Version("1.0.0")

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "key": datasets.Value("string"),
                    "audio": datasets.Audio(sampling_rate=None),
                    "dur": datasets.Value("float32"),
                    "txt": datasets.Value("string"),
                    "align": datasets.Value("string"),
                }
            ),
            supervised_keys=None,
        )

    def _split_generators(self, dl_manager):
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={
                    "jsonl_path": "eval.jsonl",
                },
            )
        ]

    def _generate_examples(self, jsonl_path):
        with open(jsonl_path, "r", encoding="utf-8") as f:
            for idx, line in enumerate(f):
                obj = json.loads(line)

                yield idx, {
                    "key": obj["key"],
                    "audio": obj["audio"], 
                    "dur": float(obj["dur"]),
                    "txt": obj["txt"],
                    # 把 struct → string
                    "align": json.dumps(obj["align"], ensure_ascii=False),
                }