22810c97b7
* use warmup_ratio as a better default than warmup steps since it's data dependent * replace remainder of warmup_steps
65 lines
1.2 KiB
YAML
65 lines
1.2 KiB
YAML
base_model: microsoft/phi-1_5
|
|
# optionally might have model_type or tokenizer_type
|
|
model_type: AutoModelForCausalLM
|
|
tokenizer_type: AutoTokenizer
|
|
# Automatically upload checkpoint and final model to HF
|
|
# hub_model_id: username/custom_model_name
|
|
|
|
load_in_8bit: false
|
|
load_in_4bit: true
|
|
|
|
datasets:
|
|
- path: garage-bAInd/Open-Platypus
|
|
type: alpaca
|
|
|
|
dataset_prepared_path:
|
|
val_set_size: 0.05
|
|
output_dir: ./outputs/phi-sft-out
|
|
|
|
sequence_len: 2048
|
|
sample_packing: true
|
|
|
|
|
|
adapter: qlora
|
|
lora_model_dir:
|
|
lora_r: 64
|
|
lora_alpha: 32
|
|
lora_dropout: 0.05
|
|
lora_target_linear: true
|
|
|
|
wandb_project:
|
|
wandb_entity:
|
|
wandb_watch:
|
|
wandb_name:
|
|
wandb_log_model:
|
|
|
|
gradient_accumulation_steps: 1
|
|
micro_batch_size: 2
|
|
num_epochs: 4
|
|
optimizer: adamw_torch_fused
|
|
adam_beta2: 0.95
|
|
adam_epsilon: 0.00001
|
|
max_grad_norm: 1.0
|
|
lr_scheduler: cosine
|
|
learning_rate: 0.000003
|
|
|
|
bf16: auto
|
|
tf32: true
|
|
|
|
gradient_checkpointing: true
|
|
gradient_checkpointing_kwargs:
|
|
use_reentrant: True
|
|
resume_from_checkpoint:
|
|
logging_steps: 1
|
|
flash_attention: true
|
|
|
|
warmup_ratio: 0.1
|
|
evals_per_epoch: 4
|
|
saves_per_epoch: 1
|
|
weight_decay: 0.1
|
|
resize_token_embeddings_to_32x: true
|
|
special_tokens:
|
|
pad_token: "<|endoftext|>"
|
|
|
|
# save_first_step: true # uncomment this to validate checkpoint saving works with your config
|