JonesLin's picture
Add validated speech alignment and natural interaction data v1
f570fd9 verified
Raw History Blame Contribute Delete
795 Bytes
policy = "Qwen/Qwen3-0.6B-Base"
policy_revision = "da87bfb608c14b7cf20ba1ce41287e8de496c0cd"
mimi_repo = "kyutai/moshiko-pytorch-bf16"
mimi_revision = "2bfc9ae6e89079a5cc7ed2a68436010d91a3d289"
device = "cuda"
audio_dim = 512
vision_dim = 768
codebooks = 8
codebook_size = 2048
audio_head_width = 256
audio_head_layers = 2
audio_head_attention_heads = 4
max_context_tokens = 2048
precision = "bfloat16"
freeze_policy = true
batch_size = 8
gradient_accumulation = 1
learning_rate = 0.0003
weight_decay = 0.01
steps = 400
save_every = 400
dataset = "codec-pilot/tts-train.jsonl"
output_dir = "runs/hf-pilot-tts"
stage = "alignment_tts"
# Opt-in: use a new run; effective batch differs from the baseline.
alignment_batching = "batched"
frozen_linear_dtype = "bfloat16"