9f824ef76a
* simplify the example configs to be more minimal and less daunting * drop empty s2_attention from example yamls
64 lines
1.2 KiB
YAML
64 lines
1.2 KiB
YAML
base_model: NousResearch/Llama-2-7b-hf
|
|
# optionally might have model_type or tokenizer_type
|
|
model_type: LlamaForCausalLM
|
|
tokenizer_type: LlamaTokenizer
|
|
# Automatically upload checkpoint and final model to HF
|
|
# hub_model_id: username/custom_model_name
|
|
strict: false
|
|
|
|
datasets:
|
|
- path: teknium/GPT4-LLM-Cleaned
|
|
type: alpaca
|
|
dataset_prepared_path: last_run_prepared
|
|
val_set_size: 0.05
|
|
output_dir: ./outputs/lisa-out
|
|
|
|
sequence_len: 4096
|
|
sample_packing: true
|
|
pad_to_sequence_len: true
|
|
|
|
adapter:
|
|
lora_model_dir:
|
|
lora_r:
|
|
lora_alpha:
|
|
lora_dropout:
|
|
lora_target_linear:
|
|
|
|
lisa_n_layers: 4
|
|
lisa_step_interval: 20
|
|
lisa_layers_attribute: model.layers
|
|
|
|
wandb_project:
|
|
wandb_entity:
|
|
wandb_watch:
|
|
wandb_name:
|
|
wandb_log_model:
|
|
|
|
gradient_accumulation_steps: 2
|
|
micro_batch_size: 1
|
|
num_epochs: 1
|
|
optimizer: adamw_bnb_8bit
|
|
lr_scheduler: cosine
|
|
learning_rate: 5e-5 # recommendation from lisa paper for 7b
|
|
|
|
bf16: auto
|
|
tf32: false
|
|
|
|
gradient_checkpointing: true
|
|
resume_from_checkpoint:
|
|
logging_steps: 1
|
|
flash_attention: true
|
|
flash_attn_cross_entropy: false
|
|
flash_attn_rms_norm: true
|
|
flash_attn_fuse_qkv: false
|
|
flash_attn_fuse_mlp: true
|
|
|
|
warmup_steps: 100
|
|
evals_per_epoch: 4
|
|
saves_per_epoch: 1
|
|
weight_decay: 0.1
|
|
special_tokens:
|
|
bos_token: "<s>"
|
|
eos_token: "</s>"
|
|
unk_token: "<unk>"
|