PUBLIC CODE anonymous supplementary material

PUBLIC IMPLEMENTATION

Code for the
standard joint pathway.

The public release keeps the full semantic-latent DiT, Qwen hidden-state readout, standalone pretraining, and joint CE + FM training logic explicit.

Release boundary

The SSL tokenizer, SSL-VAE implementation, downstream acoustic DiT, LeVo VAE decoder, checkpoints, datasets, and private infrastructure remain replaceable external components.

Complete public code

The complete public code for this demo is included in the original demopage repository under code/. The entries below link directly to the main implementation files.

MAIN ENTRY POINTS

Read the implementation by stage.

MODEL

Full semantic DiT

jointc/dit.py

Sequence-level ID and continuous-condition velocity model.

 JOINT PATH

Qwen + semantic flow

jointc/model.py

Teacher forcing, latent readout, CE + FM loss, and sampling.

 STAGE A

Standalone pretraining

jointc/warmup.py

Pretrains the semantic DiT from reference IDs and SSL-VAE targets.

 STAGE B

Joint semantic training

jointc/train.py

Loads the pretrained DiT and trains the planner, readout, and flow.

 INFERENCE

Autoregressive generation

jointc/infer.py

Samples semantic IDs and integrates the complete semantic flow.

 BOUNDARY

External adapters

jointc/adapters.py

Interfaces for the intentionally private representation and audio modules.

TRAINING STAGES

Two reproducible entry points.

Standalone semantic DiT pretraining 
PYTHONPATH=code python3 -m jointc.warmup \
  --data path/to/semantic_latents.jsonl \
  --output warmup_dit.pt
Joint semantic training 
PYTHONPATH=code python3 -m jointc.train \
  --model-name-or-path Qwen/Qwen2.5-1.5B \
  --data path/to/jointc.jsonl \
  --warmup-checkpoint warmup_dit.pt

REPOSITORY FOLDERS

What is included.

  • code/jointc/Public model and training modules
  • code/tests/CPU smoke test and shape checks
  • configs/Anonymous training templates
  • demo/Six-system qualitative audio page
  • docs/Release and anonymity checklists
← Back to audio demo