penguinfish1688/DuDE-Stage-II

DuDE-Stage-II

Committed Stage II teacher self-distillation records: 337,389 train utterances (703.460 hours of synthetic codec targets) and 128 validation utterances.

Each row is one independent utterance. The trainer samples independent utterances into two densely interleaved channels; this dataset does not assemble conversations.

Targets were generated by Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice (weights revision 0c0e3051f131929182e2c023b9537f8b1c68adfe, upstream code revision 022e286b98fbec7e1e916cb940cdf532cd9f488e). Rows preserve native text/instruction token IDs, generation settings/seeds, Ryan/Aiden voice labels, source IDs and rollout cycles. The exact same committed corpus was reused by the Stage II recovery run through step 2000.

Codec arrays have shape (frames, 16). The complete synthetic utterances were stored as codec targets, not as waveform files; they can be decoded with the matching Qwen3-TTS speech tokenizer. This release includes every committed target and excludes uncommitted generation attempts.

Attribution and license

Text source: the Parler-TTS project’s LibriTTS-R filtered speaker descriptions, clean configuration, revision 0ffbb79e07eaea6d9b191ac93a430f9264b3131d. This source is distributed under CC BY 4.0; attribution and the same license are retained for this derived dataset. DuDE added synthetic speech codec targets, native model prompts, and generation provenance. The Qwen teacher is separately licensed under Apache 2.0.

Loading

import io, json
import numpy as np
from datasets import load_dataset

ds = load_dataset("penguinfish1688/DuDE-Stage-II", "utterances", split="train", streaming=True)
row = next(iter(ds))
metadata = json.loads(row["metadata_json"])
codes = np.load(io.BytesIO(row["audio_codes_npy"]), allow_pickle=False)

audio_codes_npy contains the original NumPy payload, byte for byte (uint16, 12.5 frames/second, 16 codebooks). audio_codes_sha256 verifies each payload. frames, duration_seconds, text/xml, and conversation_id provide searchable columns; metadata_json retains the full record. Stage III duration comes from the source clock and can differ slightly from frames/12.5 due to codec rounding.

Absolute cluster paths were replaced with repository-relative assets/... paths. Other training metadata and all codec payloads are preserved. Full source annotations and voice-reference JSON live under assets/; waveform files are included for Stage III. Download those assets with huggingface_hub.snapshot_download when needed. Streaming only the Parquet records does not download the waveform assets.

To reconstruct the original training cache format after a full snapshot download:

python restore_sqlite.py --config utterances --output /path/to/cache

The helper rebases record paths to the downloaded repository and recreates each SQLite examples table with the original codec bytes. It does not restore generation worker cursors. It refuses to overwrite existing cache files.

manifest.json records verified row counts, durations, Parquet checksums and record digests. files.jsonl lists the upload files and sizes; hashes are included for generated metadata and Parquet files. Original audio files are uploaded without re-encoding. The release contains datasets, not model checkpoints.