Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

PolyPlex PersonaPlex Pairs

English full-duplex conversations generated by two copies of nvidia/personaplex-7b-v1 talking to each other. PolyPlex uses them to keep its English duplex behavior close to the original PersonaPlex: the frames are fed to PersonaPlex and to the trained model with teacher forcing, and a KL loss is taken on the text and audio outputs.

How the conversations were made

  • Two PersonaPlex models face each other and generate frame by frame. One side's user input at frame t is the other side's assistant audio at frame t-2 (the first two frames are codec silence), because the depth decoder fills a frame's acoustic codebooks one step later.
  • Personas come in pairs: one side uses an example persona from the PersonaPlex README, the other its counterpart (a customer, a student, a conversation partner). A file uses one persona pair, cycling through 8 pairs.
  • Sampling: temperature 0.8, top_k 250. Each conversation is 750 frames (60 s) after the persona prefix.
  • Only the text tokens and both sides' Mimi codes are stored. No waveform is included.
  • Training and evaluation conversations are split by seed range.

Files

Folder Files Seeds Conversations Rows Use
train/ 125 0 to 124 2,000 4,000 Training
train-2/ 75 2000 to 2074 1,200 2,400 Training
eval/ 3 100000 to 100002 64 128 Evaluation only, never trained on

Each selfplay-<seed>.pt file is a list saved with torch.save. One conversation gives two rows, one from each side's point of view. Every row is a dict:

Key Type Meaning
persona str The persona of this row's assistant side
partner str The persona of the other side
seed int Seed of the file
conversation int Index of the conversation in the file
prefix_frames int Number of persona prefix frames before the conversation
input_ids LongTensor (frames,) The assistant's text tokens (PersonaPlex vocabulary)
assistant_audio_codes LongTensor (8, frames) The assistant's Mimi codes
user_audio_codes LongTensor (8, frames) The other side's Mimi codes, as this side heard them

Load with torch.load(path, weights_only=False). The files are pickles, so load only files you trust.

License

The conversations are outputs of nvidia/personaplex-7b-v1 and are released under the NVIDIA Open Model License.

Downloads last month
65