File size: 1,642 Bytes
aec27cf 432dcad aec27cf |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 |
data:
prompt: llama3_formal
train: ../data/susgen/FINAL/PER_3500/FINAL_PER3500_30k.json
val: null
val_split_ratio: 0.005
device: cuda
instruct_mask: true
local_rank: 0
model:
acceleration: null
int4_config:
bnb_4bit_compute_dtype: bfloat16
bnb_4bit_quant_type: nf4
bnb_4bit_use_double_quant: true
load_in_4bit: true
load_in_8bit: false
int8_config:
load_in_4bit: false
load_in_8bit: true
lora:
bias: none
inference_mode: false
lora_alpha: 32
lora_dropout: 0.1
r: 16
target_modules:
- q_proj
- k_proj
- v_proj
- o_proj
- gate_proj
- up_proj
- down_proj
- lm_head
task_type: CAUSAL_LM
lora_path: false
model_path: ../ckpts/Meta-Llama-3-8B-Instruct
quantization: int4
seed: 2024
show_config: false
use_lora: true
name: 30k-Llama3-8B-Instruct
output_dir: ../results/
tokenizer:
add_bos_token: true
add_eos_token: false
add_prefix_space: false
encode:
max_length: 2048
return_tensors: pt
truncation: true
model_max_length: 2048
padding_side: left
pretrained_model_name_or_path: ../ckpts/Meta-Llama-3-8B-Instruct
truncation_side: right
use_fast: true
trainer: NewTrainer
training:
bf16: true
deepspeed: ./configs/ds_configs/ds_config_stage_2.json
gradient_accumulation_steps: 16
learning_rate: 2.0e-05
logging_steps: 1
lr_scheduler_type: cosine
num_train_epochs: 3
optim: paged_adamw_32bit
per_device_train_batch_size: 16
remove_unused_columns: false
report_to: wandb
resume_from_checkpoint: true
save_steps: 20
save_strategy: steps
warmup_steps: 100
weight_decay: 0.01
|