76 lines
1.7 KiB
YAML
76 lines
1.7 KiB
YAML
seed: 1
|
|
exp_name: llama3-8B-sft
|
|
train_datasets:
|
|
- yunjae-won/Qwen3-30B-MagpieLM-SFT-Outputs-v0.1-shard0
|
|
test_datasets:
|
|
- yunjae-won/Qwen3-30B-MagpieLM-SFT-Outputs-v0.1-shard0
|
|
debug: false
|
|
wandb:
|
|
enabled: true
|
|
entity: null
|
|
project: KD
|
|
cache_dir: .cache/
|
|
local_run_dir: .cache//llama3-8B-sft
|
|
do_first_eval: true
|
|
minimum_log_interval_secs: 1.0
|
|
intermediate_checkpoints: false
|
|
trainer: BasicTrainer
|
|
template_tokens: []
|
|
lr: 5.0e-06
|
|
n_epochs: 1
|
|
n_examples: null
|
|
n_eval_examples: 512
|
|
eval_every: 19968
|
|
save_every: 5120
|
|
step_scheduler_with_optimizer: false
|
|
optimizer: RMSprop
|
|
weight_decay: 0
|
|
beta1: 0.9
|
|
beta2: 0.999
|
|
eps: 1.0e-05
|
|
warmup: 0.1
|
|
cache_reference_logprobs: false
|
|
load_reference_logprobs: null
|
|
humanline: false
|
|
log_epsilon_P: -1.0
|
|
log_epsilon_R: 1.5
|
|
online: false
|
|
frac_unique_desirable: 1.0
|
|
frac_unique_undesirable: 1.0
|
|
model:
|
|
name_or_path: meta-llama/Llama-3.1-8B
|
|
tokenizer_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
|
|
load_from: null
|
|
from_checkpoint: null
|
|
block_name: LlamaDecoderLayer
|
|
policy_dtype: bfloat16
|
|
reference_dtype: bfloat16
|
|
max_grad_norm: 10.0
|
|
v_head_max_grad_norm: 0.1
|
|
max_length: 4096
|
|
max_prompt_length: 2048
|
|
activation_checkpointing: false
|
|
batch_size: 256
|
|
microbatch_size: 2.0
|
|
gradient_accumulation_steps: 32
|
|
eval_batch_size: 256
|
|
eval_microbatch_size: 2.0
|
|
attn_implementation: flash_attention_2
|
|
use_peft: false
|
|
load_lora_from: null
|
|
peft:
|
|
lora_r: 64
|
|
lora_alpha: 256
|
|
lora_dropout: 0.05
|
|
target_modules: all-linear
|
|
reward_model:
|
|
path: null
|
|
model_class: AutoModelForBradleyTerry
|
|
dtype: float32
|
|
attn_implementation: flash_attention_2
|
|
loss:
|
|
trainer: SFTTrainer
|
|
dataloader: SFTDataLoader
|
|
sync_reference: false
|
|
num_epochs: 1
|