llama2-7b-comunism-br-pt-qlora.yml

base_model: 'Nous-Hermes-llama-2-7b' 
base_model_config: 'Nous-Hermes-llama-2-7b'
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
is_llama_derived_model: true

trust_remote_code: true
# use_fast option for tokenizer loading from_pretrained, default to True
tokenizer_use_fast: True
# Whether to use the legacy tokenizer setting, defaults to True
tokenizer_legacy: False

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: ./split-todos-artigos-limpos-v7.json
    type: completion
dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./comunismo-br-hermes-7b-llama2-lm-head-cleaner

adapter: qlora
lora_model_dir:

sequence_len: 4096
#max_packed_sequence_len: 4096
lora_r: 128
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
  - q_proj
  - v_proj
  - k_proj
  - o_proj
  - gate_proj
  - down_proj
  - up_proj
  - lm_head
  - wte
  - embed_tokens
#lora_target_modules:
lora_modules_to_save:
  - wte
  - embed_tokens
  - lm_head
#lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: comunismo-br-hermes-7b-llama2-lm-head
wandb_watch:
wandb_run_id:
wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 2
num_epochs: 1
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 0.000015

train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
#xformers_attention: true
flash_attention: true

#max_steps: 5
warmup_steps: 100
#eval_steps: 50
save_steps: 500
save_total_limit: 10
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
#special_tokens:
#  bos_token: "<s>"
#  eos_token: "</s>"
#  unk_token: "<unk>"
#  pad_token: "<unk>"