example_axolotl_config_for_dpo_on_tw_chatbot_arena_data

dpo.yaml

base_model: yentinglin/Llama-3-Taiwan-8B-Instruct

load_in_8bit: false
load_in_4bit: false
strict: false

hub_model_id: yentinglin/Llama-3-Taiwan-8B-Instruct-DPO
hub_strategy: end
wandb_name: 8b dpo

dataset_processes: 1
rl: dpo
datasets:
  - path: aigrant/tw_chatbot_arena
    name: argilla
    split: train
    type: llama3.argilla_chat

chat_template: llama3

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./output/tw-llm/8b/dpo/

sequence_len: 8000
pad_to_sequence_len: true

gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 4
optimizer: adamw_8bit
lr_scheduler: cosine
learning_rate: 5e-5

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 100
evals_per_epoch: 0
eval_table_size:
saves_per_epoch: 1
save_total_limit: 100
save_steps:
debug:
deepspeed: deepspeed_configs/zero3_bf16.json
weight_decay: 0.05
fsdp:
fsdp_config: