Training and export scripts for Luciole Audio: speech/audio-augmented LLMs (SALM)
built on the Luciole models (1B, 8B, 23B) with NVIDIA NeMo speechlm2.
| Folder | Content |
|---|---|
data/ |
Dataset conversion to NeMo/Lhotse manifests, data generation, templating, TTS |
training/ |
SALM training (salm_train.py), Hydra configs (conf/), Slurm launcher |
export/ |
Export of NeMo checkpoints to Hugging Face format (to_hf.py) |
utils/ |
Shared helpers (audio, inference, metrics) |
assets/ |
Model/dataset cards, plotting and upload scripts |
demo/ |
WebRTC demo app |
install/ |
Environment installation scripts (Jean-Zay) |
Data preparation scripts (data/) import luciole_audio.*, so the repo must be installed:
pip install -e ".[data]" # or [generation], [processing], [templating], [all-data], ...The tts extra conflicts with the others (different transformers pin) and must be
installed in a separate environment.
Training (submits training/run_train.slurm):
python training/slurm_launcher.py --config run/luciole-8b --gpus 4 --time 20:00:00Hugging Face export:
sbatch export/run_hf_export.slurmModels are evaluated with our fork of AudioBench (results on the leaderboard).
AGPL v3, see LICENSE.