69 lines
1.4 KiB
YAML
69 lines
1.4 KiB
YAML
|
base_model: winglian/Llama-2-3b-hf
|
||
|
|
base_model_config: winglian/Llama-2-3b-hf
|
||
|
|
model_type: LlamaForCausalLM
|
||
|
|
tokenizer_type: LlamaTokenizer
|
||
|
|
load_in_8bit: false
|
||
|
|
load_in_4bit: false
|
||
|
|
hub_model_id: winglian/llama-2-3b-cpt
|
||
|
|
strict: false
|
||
|
|
# pretraining_dataset: togethercomputer/RedPajama-Data-1T-Sample
|
||
|
|
push_dataset_to_hub:
|
||
|
|
datasets:
|
||
|
|
- path: togethercomputer/RedPajama-Data-1T-Sample
|
||
|
|
type: completion
|
||
|
|
dataset_prepared_path: last_run_prepared
|
||
|
|
val_set_size: 0.0
|
||
|
|
adapter:
|
||
|
|
lora_model_dir:
|
||
|
|
sequence_len: 4096
|
||
|
|
sample_packing:
|
||
|
|
lora_r:
|
||
|
|
lora_alpha:
|
||
|
|
lora_dropout:
|
||
|
|
lora_target_modules:
|
||
|
|
lora_target_linear:
|
||
|
|
lora_fan_in_fan_out:
|
||
|
|
wandb_project: llama-2-3b-cpt
|
||
|
|
wandb_entity:
|
||
|
|
wandb_watch:
|
||
|
|
wandb_run_id:
|
||
|
|
wandb_log_model:
|
||
|
|
output_dir: ./llama-2-3b-out
|
||
|
|
# max_steps: 4400
|
||
|
|
gradient_accumulation_steps: 3
|
||
|
|
micro_batch_size: 8
|
||
|
|
num_epochs: 1
|
||
|
|
optimizer: adamw_torch
|
||
|
|
adam_beta2: 0.95
|
||
|
|
torchdistx_path:
|
||
|
|
lr_scheduler: cosine
|
||
|
|
learning_rate: 0.0001
|
||
|
|
train_on_inputs: false
|
||
|
|
group_by_length: false
|
||
|
|
float16: true
|
||
|
|
bf16: truc
|
||
|
|
fp16: false
|
||
|
|
tf32: true
|
||
|
|
gradient_checkpointing: true
|
||
|
|
early_stopping_patience:
|
||
|
|
resume_from_checkpoint:
|
||
|
|
local_rank:
|
||
|
|
logging_steps: 1
|
||
|
|
xformers_attention:
|
||
|
|
flash_attention: true
|
||
|
|
gptq_groupsize:
|
||
|
|
gptq_model_v1:
|
||
|
|
warmup_steps: 100
|
||
|
|
eval_steps:
|
||
|
|
save_steps: 100
|
||
|
|
save_total_limit: 6
|
||
|
|
debug:
|
||
|
|
deepspeed:
|
||
|
|
weight_decay: 0.1
|
||
|
|
fsdp:
|
||
|
|
fsdp_config:
|
||
|
|
special_tokens:
|
||
|
|
bos_token: "<s>"
|
||
|
|
eos_token: "</s>"
|
||
|
|
unk_token: "<unk>"
|