WTForbes/hak

Consolidated hak dataset repo staged from referenced corpus JSONL rows only.

Status

• corpus_jsonl: staged
• root_corpus_jsonl: completed
• subset_corpus_jsonl: completed
• manifests: staged
• manifests_upload: completed
• audio_parquet: completed
• upload: completed
• upload_policy: referenced audio only; never upload whole source directories

Layout

corpus_dir_jsonl/{train,train_stage1,eval}.jsonl
subsets/<subset>/
  corpus_dir_jsonl/{train,train_stage1,eval}.jsonl
  manifests/audio_manifest.tsv
  audio_parquet/part-xxxxx.parquet

corpus_dir_jsonl/ contains the full un-split corpus for hak. subsets/ keeps the source-set grouping used for selected corpus files, manifests, and audio parquet shards.

Rules

• Preserve original_audio_filepath in every corpus row.
• Rewrite audio_filepath to subsets/<subset>/audio/<source_set>/<sha shard>/<audio_id>.<stored_ext>.
• Convert WAV to FLAC before writing Parquet.
• Keep existing FLAC/MP3/Opus/Ogg codecs unless explicitly changed later.
• Use deterministic audio_id and SHA shard paths to avoid basename collisions.

Subsets

• hak_yijiazhizhu from hak_hak__yijiazhizhu
• hakka_processed_2023_year_hakka_collection from hak_hakka_processed__2023_year_hakka_collection
• hakka_processed_2024_year_hakka_collection from hak_hakka_processed__2024_year_hakka_collection
• hakka_processed_2025_year_hakka_collection from hak_hakka_processed__2025_year_hakka_collection
• hakka_processed_2026_year_hakka_collection from hak_hakka_processed__2026_year_hakka_collection
• hakka_processed_audio_fixed_25146 from hak_hakka_processed__audio_fixed_25146
• hakka_processed_eattogther from hak_hakka_processed__eattogther
• hakka_processed_show_hou from hak_hakka_processed__show_hou
• hakka_processed_singing from hak_hakka_processed__singing
• hakka_processed_whotoeat from hak_hakka_processed__whotoeat