58 lines
1.3 KiB
YAML
58 lines
1.3 KiB
YAML
base_model: ibm-granite/granite-speech-3.3-2b
|
|
# Remove model_type to let Axolotl auto-detect the correct model type
|
|
# model_type: GraniteSpeechForConditionalGeneration
|
|
# Enable trust_remote_code to use the model's custom code
|
|
trust_remote_code: true
|
|
# Mark as multimodal since this is a speech model
|
|
is_multimodal: true
|
|
|
|
hub_model_id: syvai/gsp
|
|
|
|
plugins:
|
|
- axolotl.integrations.liger.LigerPlugin
|
|
liger_rope: true
|
|
liger_rms_norm: true
|
|
liger_glu_activation: true
|
|
liger_fused_linear_cross_entropy: true
|
|
|
|
datasets:
|
|
- path: syvai/coral-tts-asr
|
|
type: # leave empty to load pre-tokenized
|
|
dataset_prepared_path: last_run_prepared
|
|
val_set_size: 0.02
|
|
output_dir: ./outputs/out
|
|
eval_sample_packing: False
|
|
|
|
sequence_len: 8192
|
|
sample_packing: true
|
|
pad_to_sequence_len: true
|
|
|
|
wandb_project:
|
|
wandb_entity:
|
|
wandb_watch:
|
|
wandb_name:
|
|
wandb_log_model:
|
|
|
|
gradient_accumulation_steps: 16
|
|
micro_batch_size: 1
|
|
num_epochs: 3
|
|
optimizer: adamw_torch_fused
|
|
lr_scheduler: cosine
|
|
learning_rate: 2e-5
|
|
|
|
bf16: auto
|
|
tf32: false
|
|
|
|
gradient_checkpointing: true
|
|
gradient_checkpointing_kwargs:
|
|
use_reentrant: false
|
|
resume_from_checkpoint:
|
|
logging_steps: 1
|
|
flash_attention: true
|
|
|
|
warmup_steps: 20
|
|
evals_per_epoch: 5
|
|
saves_per_epoch: 5
|
|
weight_decay: 0.05
|
|
|
|
#save_first_step: true # uncomment this to validate checkpoint saving works with your config |