Voxtral Creole training scripts

Training and eval scripts for chendren/voxtral-creole-lora-v3. This repo is code and test wavs, not the adapter.

train.py, train_v2.py, train_v3.py Hugging Face Jobs / CUDA LoRA
benchmark.py, blind_test.py, demo_inference.py
test-audio/ short clips used in write-ups
synthesize_domains.py domain-prompt experiments

Released adapter: PEFT LoRA on mistralai/Voxtral-Mini-3B-2507. CMU Haitian Creole Speech held-out: WER 9.7, CER 3.1, 192/500 exact. Base model: WER 101.4, 0 exact.

Private domain audio stays at chendren/creole-domain-audio (gated). Public WER is the CMU split.

Demo Space (examples, no GPU): chendren/voxtral-creole-demo

# adapter inference lives on the model repo
# this repo is the training tree
python benchmark.py

Downloads last month: -; Downloads are not tracked for this model. How to track

Collection including chendren/voxtral-creole-training

Haitian Creole STT

Collection

Voxtral Mini 3B LoRA for Haitian Creole. Public WER is CMU held-out 9.7. • 3 items • Updated 24 days ago