64 lines
1.2 KiB
YAML
64 lines
1.2 KiB
YAML
base_model: meta-llama/Llama-3.2-1B
|
|
# Automatically upload checkpoint and final model to HF
|
|
# hub_model_id: username/custom_model_name
|
|
|
|
load_in_8bit: false
|
|
load_in_4bit: true
|
|
|
|
rl: kto
|
|
rl_beta: 0.5
|
|
kto_desirable_weight: 0.2
|
|
|
|
datasets:
|
|
- path: argilla/ultrafeedback-binarized-preferences-cleaned-kto
|
|
type: llama3.ultra
|
|
split: train
|
|
dataset_prepared_path: last_run_prepared
|
|
val_set_size: 0.0
|
|
output_dir: ./outputs/qlora-out
|
|
|
|
remove_unused_columns: false
|
|
|
|
adapter: qlora
|
|
lora_model_dir:
|
|
|
|
sequence_len: 2048
|
|
sample_packing: false # not supported with kto
|
|
eval_sample_packing: false
|
|
pad_to_sequence_len: false
|
|
|
|
lora_r: 32
|
|
lora_alpha: 64
|
|
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: 1
|
|
optimizer: adamw_8bit
|
|
lr_scheduler: cosine
|
|
learning_rate: 0.0002
|
|
|
|
bf16: auto
|
|
tf32: true
|
|
|
|
gradient_checkpointing: true
|
|
gradient_checkpointing_kwargs:
|
|
use_reentrant: false
|
|
resume_from_checkpoint:
|
|
logging_steps: 1
|
|
flash_attention: true
|
|
|
|
warmup_steps: 20
|
|
evals_per_epoch: 4
|
|
saves_per_epoch: 1
|
|
weight_decay: 0.0
|
|
special_tokens:
|
|
pad_token: "<|end_of_text|>"
|