training config
Browse files- train_config.yaml +19 -0
train_config.yaml
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
data: {train_data: "/home/coherent_pro_full", eval_data: "", shuffle: true}
|
| 2 |
+
moshi_paths: {hf_repo_id: "kyutai/moshiko-pytorch-bf16"}
|
| 3 |
+
full_finetuning: false
|
| 4 |
+
lora: {enable: true, rank: 64, scaling: 2., ft_embed: false}
|
| 5 |
+
first_codebook_weight_multiplier: 100.
|
| 6 |
+
text_padding_weight: .5
|
| 7 |
+
duration_sec: 20
|
| 8 |
+
batch_size: 1
|
| 9 |
+
max_steps: 1500
|
| 10 |
+
gradient_checkpointing: true
|
| 11 |
+
optim: {lr: 2.e-5, weight_decay: 0.1, pct_start: 0.05}
|
| 12 |
+
seed: 0
|
| 13 |
+
log_freq: 50
|
| 14 |
+
eval_freq: 20000
|
| 15 |
+
do_eval: false
|
| 16 |
+
do_ckpt: true
|
| 17 |
+
ckpt_freq: 1500
|
| 18 |
+
save_adapters: true
|
| 19 |
+
run_dir: "/home/moshi_run_pro_full"
|