split train.parquet into 6 shards (<300MB each) so the HF dataset viewer works; README: shard layout + link DaoZang-Online space

single-split dataset: drop files.parquet (auto-inference requires standard split names and unified schema)

docs: embeddings are L2-normalized

DaoZang dataset: 285,117 chunks with bge-m3 embeddings + 3,152 files + uncompressed markdowns