This model is based on Whisper Large v3 Turbo, and retrained for Catalan with Mozilla Common Voice Dataset 26.0.
NOTE: If you want to use with CTranslate2 / Faster Whisper, check the converted model here.
Checkpoint: 131025 (2 epochs)
lr: 1e-4
precision: fp16
lora = LoraConfig(
r=64,
lora_alpha=128,
target_modules=["q_proj", "k_proj", "v_proj", "out_proj"],
lora_dropout=0.05,
bias="none",
rsLoRA=True,
)