Upload code/configs/ablation_nat_speaker.yaml with huggingface_hub
Browse files
code/configs/ablation_nat_speaker.yaml
ADDED
|
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Ablation D: Naturalness + Speaker Similarity
|
| 2 |
+
# Tests the effect of adding speaker similarity alone (no quality MLP, no centroid)
|
| 3 |
+
# From LaionBox v0.1-wip, same settings as nat_only_2ep
|
| 4 |
+
|
| 5 |
+
# ββ Data source ββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 6 |
+
preprocessed_dir: "./finetune_data_combined"
|
| 7 |
+
index_file: "./finetune_data_combined/index_mixed.json"
|
| 8 |
+
|
| 9 |
+
bucket_weights:
|
| 10 |
+
dramabox: 0.4
|
| 11 |
+
emolia: 0.2
|
| 12 |
+
podcast: 0.4
|
| 13 |
+
|
| 14 |
+
podcast_epochs: 2
|
| 15 |
+
|
| 16 |
+
# ββ DramaBox model paths ββββββββββββββββββββββββββββββββββββββββββββββ
|
| 17 |
+
dramabox_dir: "/home/deployer/laion/DramaBox"
|
| 18 |
+
checkpoint: "/home/deployer/laion/DramaBox/models/ltx-2.3-22b-dev-audio-only-v13-merged.safetensors"
|
| 19 |
+
full_checkpoint: "/home/deployer/laion/DramaBox/models/ltx-2.3-22b-dev.safetensors"
|
| 20 |
+
gemma_root: "/home/deployer/.cache/dramabox/models--unsloth--gemma-3-12b-it-bnb-4bit/snapshots/826e729dbaeea4ecb143738eed2bcf3539ebf7bf"
|
| 21 |
+
base_model: "dev"
|
| 22 |
+
|
| 23 |
+
# ββ LoRA ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 24 |
+
lora_rank: 128
|
| 25 |
+
lora_alpha: 128
|
| 26 |
+
lora_dropout: 0.05
|
| 27 |
+
resume_lora: "/home/deployer/.cache/huggingface/hub/models--laion--laionbox-v0.1-wip/snapshots/66176d2a653a013a7b71c1ccb7a7a4d4cf514b0d/lora_epoch5.safetensors"
|
| 28 |
+
|
| 29 |
+
# ββ Training ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 30 |
+
lr: 2.0e-5
|
| 31 |
+
lr_scheduler: "cosine"
|
| 32 |
+
warmup_steps: 40
|
| 33 |
+
batch_size: 1
|
| 34 |
+
grad_accum: 32
|
| 35 |
+
max_grad_norm: 1.0
|
| 36 |
+
seed: 42
|
| 37 |
+
expand_all_modes: true
|
| 38 |
+
max_ref_tokens: 250
|
| 39 |
+
text_dropout: 0.1
|
| 40 |
+
|
| 41 |
+
# ββ Auxiliary rewards βββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 42 |
+
clap_loss: true
|
| 43 |
+
differentiable_reward: false
|
| 44 |
+
clap_model: "gijs/voiceclap-lco-7b-lora"
|
| 45 |
+
clap_quantize: "int4"
|
| 46 |
+
aux_sigma_max: 0.4
|
| 47 |
+
aux_target_ratio: 6.0
|
| 48 |
+
coeff_cap: 10.0
|
| 49 |
+
no_centroid: true
|
| 50 |
+
no_speaker_sim: false # ENABLED: speaker similarity (WavLM-SV)
|
| 51 |
+
no_quality_mlp: true
|
| 52 |
+
positive_text: "Realistic, genuine, spontaneous, authentic, sensual, natural voice with all imperfections and organic microdistractions a natural situation brings with it"
|
| 53 |
+
negative_text: "distorted, unnatural, robotic, distortion"
|
| 54 |
+
|
| 55 |
+
# ββ Checkpoint management ββββββββββββββββββββββββββββββββββββββββββββ
|
| 56 |
+
output_dir: "./finetune_output/ablation_nat_speaker"
|
| 57 |
+
save_every: 10
|
| 58 |
+
keep_last_n: 3
|
| 59 |
+
log_every: 10
|
| 60 |
+
monitor_port: 8771
|
| 61 |
+
|
| 62 |
+
val_samples: 0
|
| 63 |
+
|
| 64 |
+
no_save_state: true
|