scripts/dataset/README.md

Dataset preparation

LTX's trainer takes a dataset; it does not help you build one. For an audio LoRA the dataset is the challange, the model learns a transformation, so every item must be the same performance twice, once dry and once through a real acoustic treatment, aligned and gain-matched. I also used two iterations of each dry sound source so model can learn on variety. These are the scripts that built the AKUSPACE set (266 pairs, 14 sources × 19 cells).

Set AKUSPACE_DATASET to your dataset root, or run from the repo and the relative default applies.

Script	Does
build_v5_rooms.py	dry/wet room pairs at three levels per room
build_v5_outdoor.py	outdoor beds; projects the dry out before rebuilding
build_v5_sfx.py	Eurorack modular-delay pairs
build_eurorack_dry.py	assembles the dry side fed through the hardware
make_manifest_v5.py	writes the training CSV with captions and splits
verify_dataset.py	checks every pair: length, rate, ceiling, alignment
check_limiting.py	catches limiting/clipping introduced by a render chain
check_clips.py	duration and sample-rate conformance
wetdry.py	measures the wet/dry relationship of a rendered pair
slice_ableton_stems.py	cuts long recordings into per-source clips

The rules these encode

Each one cost real time to learn.

Never normalise dry and wet independently. Apply the same gain to both, or the pair no longer describes a transformation and the model learns loudness instead of acoustics.

Normalise by active loudness, not peak. Sparse material — handclaps especially — reads ~13 dB quiet on peak-based normalisation and ends up mismatched against sustained sources.

An ambience bed is not a reverb tail. A tail is part of the signal and scales cleanly; a bed is a separate recording, so boosting it amplifies its own noise floor. Scale beds down only, and only a little. This is why the outdoor cells ship two levels while rooms and SFX ship three.

Check the dry fit before trusting any ambience measurement. R = wet − dry assumes the render carries the dry at exactly 1.00×. When it does not, a negative ghost of the dry hides inside R and dominates every derived number — which produced months of misleading ambience statistics here before verify_dataset.py caught it. Confirm fit ≈ 1.000 first.

The trigger word belongs in preprocessing, not the manifest. It is prepended to training captions by the preprocessor and must appear in inference prompts; storing it in the CSV double-applies it.