mirror of
https://github.com/hanzoai/tts.git
synced 2026-11-02 12:16:08 +00:00
26 lines
646 B
YAML
26 lines
646 B
YAML
# Model
|
|
model_name: "LiquidAI/LFM2.5-350M"
|
|
tokenizer_name: "LiquidAI/LFM2-350M" # LFM2.5 uses same tokenizer base
|
|
|
|
# Codec
|
|
codec_type: "mimi"
|
|
num_codebooks: 8 # mimi: 1-32 (default 8)
|
|
|
|
# Training Args
|
|
epochs: 1
|
|
batch_size: 16
|
|
number_processes: 8
|
|
pad_token: 64407
|
|
save_steps: 12000
|
|
learning_rate: 5.0e-5
|
|
ratio: "2:1"
|
|
|
|
# Datasets — local path or HuggingFace repo
|
|
text_QA_dataset: "/scratch/kadirnar/emilia-subset-mimi-lfm2_5"
|
|
TTS_dataset: "/scratch/kadirnar/emilia-subset-mimi-lfm2_5"
|
|
|
|
# Naming and paths
|
|
save_folder: "/scratch/kadirnar/checkpoints/lfm2_5-pretrain"
|
|
project_name: "lfm2_5-tts-mimi"
|
|
run_name: "lfm2_5-tts-mimi-pretrain"
|