📦 MOSS Dataset Prep

This is the heavy-lifting step in the MOSS-TTS training chain. Staging put your clips on disk and Dataset Rows paired them with text; Dataset Prep turns that raw pairing into the encoded, model-ready TRAINING_DATASET the trainer consumes. It runs your audio through the MOSS codec/tokenizer, splits off a validation set, and caches the result so you don't re-encode every time you tweak a hyperparameter.

If you've trained image LoRAs, this is the equivalent of the caching/latents step - the slow, one-time preprocessing that makes the actual training loop fast. Get it right once and every subsequent training run reuses it.

How it works

You connect the MOSS engine (so it knows which tokenizer/codec to use) and point it at your dataset source. It encodes each clip into the token representation MOSS trains on, optionally encodes reference audio, carves out a validation split for measuring progress, and writes a reusable prep cache. Because it caches, flipping a training setting later doesn't force a full re-encode - a real time-saver on anything but a tiny dataset.

The inputs and outputs that matter

• TTS_engine (TTS_ENGINE) - the MOSS engine node. This tells Prep which model family it's encoding for. Required.
• dataset_source (STRING) - where the dataset lives. In the standard flow this ties back to what Clip Staging / Dataset Rows produced.
• model_name (default MyMossLoRA) - the name your trained adapter will carry. Set it per project.
• validation_split (default 0.05) and split_seed (default 42) - hold out 5% of clips to validate against, with a fixed seed so the split is reproducible. 5% is a fine default for a small dataset.
• reuse_existing (default on) - reuse the prep cache if one already exists. Leave it on; that's the whole point of caching. Turn it off only when you've changed the source data and need a clean re-encode.
• encode_reference_audio (default on), prep_batch_size (default 8), n_vq (default 0) - encoding details. Lower prep_batch_size if prep runs out of memory; otherwise the defaults are sensible.

Outputs: training_dataset (TRAINING_DATASET) - feed this to the Model Training node alongside a Training Config - and dataset_info (STRING), a report of clip counts and the train/validation split.

How to install it

Part of TTS Audio Suite. ComfyUI Manager: search TTS Audio Suite, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/diodiogod/TTS-Audio-Suite.git
cd TTS-Audio-Suite
python install.py

then restart. On Linux, install portaudio19-dev and libsamplerate0-dev first. Prep needs the MOSS tokenizer/codec loaded (via the engine), which auto-downloads into ComfyUI/models/TTS/moss_tts/ on first use - that's a multi-GB download, so the first prep is slow.

Common issues & troubleshooting

Prep runs out of VRAM/RAM. Encoding is the memory-hungry part. Drop prep_batch_size (8 → 4 → 2). MOSS training in general is not a light job - this pipeline targets the 8B model, so a small GPU will struggle even before the training loop starts.

It re-encodes everything every run. Check reuse_existing is on and that model_name / dataset_source haven't changed - the cache is keyed to those. Renaming the model invalidates the cache.

Validation loss looks weird / non-existent. With a tiny dataset a 5% split might be just one or two clips. Nudge validation_split up a little (or accept that validation is noisy on small sets) - it's a health signal, not gospel.

"Wrong engine" errors. The TTS_engine input must be a MOSS engine. Wiring in a different engine's config won't prep correctly - this pipeline is MOSS-specific.