base_model: huggyllama/llama-13b
base_model_config: huggyllama/llama-13b
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
load_in_8bit: false
datasets:
  - path: vicgalle/alpaca-gpt4
    type: alpaca
#  - path: anon8231489123/ShareGPT_Vicuna_unfiltered
#    data_files: ShareGPT_V3_unfiltered_cleaned_split_no_imsorry.json
#    type: sharegpt
#  - path: ehartford/WizardLM_alpaca_evol_instruct_70k_unfiltered
#    type: alpaca
dataset_prepared_path: data/last_run_prepared
val_set_size: 0.04
adapter:
lora_model_dir:
sequence_len: 2048
max_packed_sequence_len: 2048
lora_r: 8
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
  - q_proj
  - v_proj
#  - k_proj
#  - o_proj
lora_fan_in_fan_out: false
wandb_project:
wandb_watch:
wandb_run_id:
wandb_log_model: checkpoint
output_dir: ./wizard-lm-out
batch_size: 128
micro_batch_size: 1
num_epochs: 2
warmup_steps: 46
logging_steps:
learning_rate: 0.00003
optimizer: adamw_torch
torchdistx_path:
lr_scheduler: one_cycle
log_sweep_min_lr: 2e-6
log_sweep_max_lr: 1e-4
train_on_inputs: false
group_by_length: false
bf16: true
tf32: true
gradient_checkpointing:
early_stopping_patience:
resume_from_checkpoint:
auto_resume_from_checkpoints:
local_rank:
load_4bit:
xformers_attention:
flash_attention: true
gptq_groupsize:
gptq_model_v1:
save_steps:
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
  fsdp_transformer_layer_cls_to_wrap:
  fsdp_min_num_params: 2000
  fsdp_backward_prefetch:
    - backward_pre
  limit_all_gathers: false
special_tokens:
  pad_token: "[PAD]"
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"