axolotl/pretrain_test.yaml

base_model: theblackcat102/whale-v3-base-merged
# optionally might have model_type or tokenizer_type
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: true
# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name

load_in_8bit: false
load_in_4bit: true
strict: false

adapter: qlora
lora_r: 380
lora_alpha: 380
lora_target_modules: [down_proj,gate_proj,up_proj]
lora_target_linear: false
peft_use_rslora: true

max_steps: 100000
datasets:
  - path: cerebras/SlimPajama-627B
    data_files:
      - train/chunk1/example_train_0.jsonl.zst
      - train/chunk1/example_train_1.jsonl.zst
      - train/chunk1/example_train_2.jsonl.zst
      - train/chunk1/example_train_3.jsonl.zst
      - train/chunk1/example_train_4.jsonl.zst
      - train/chunk1/example_train_5.jsonl.zst
      - train/chunk1/example_train_6.jsonl.zst
      - train/chunk1/example_train_7.jsonl.zst
      - train/chunk1/example_train_8.jsonl.zst
      - train/chunk1/example_train_9.jsonl.zst
      - train/chunk1/example_train_10.jsonl.zst
      - train/chunk1/example_train_11.jsonl.zst
      - train/chunk1/example_train_12.jsonl.zst
      - train/chunk1/example_train_13.jsonl.zst
      - train/chunk1/example_train_14.jsonl.zst
      - train/chunk1/example_train_15.jsonl.zst
      - train/chunk1/example_train_16.jsonl.zst
      - train/chunk1/example_train_17.jsonl.zst
      - train/chunk1/example_train_18.jsonl.zst
      - train/chunk1/example_train_19.jsonl.zst
      - train/chunk1/example_train_20.jsonl.zst
      - train/chunk1/example_train_21.jsonl.zst
      - train/chunk1/example_train_22.jsonl.zst
      - train/chunk1/example_train_23.jsonl.zst
      - train/chunk1/example_train_24.jsonl.zst
      - train/chunk1/example_train_25.jsonl.zst
      - train/chunk1/example_train_26.jsonl.zst
      - train/chunk1/example_train_27.jsonl.zst
      - train/chunk1/example_train_28.jsonl.zst
      - train/chunk1/example_train_29.jsonl.zst
      - train/chunk1/example_train_30.jsonl.zst
      - train/chunk1/example_train_31.jsonl.zst
      - train/chunk1/example_train_32.jsonl.zst
      - train/chunk1/example_train_33.jsonl.zst
      - train/chunk1/example_train_34.jsonl.zst
      - train/chunk1/example_train_35.jsonl.zst
      - train/chunk1/example_train_36.jsonl.zst
      - train/chunk1/example_train_37.jsonl.zst
      - train/chunk1/example_train_38.jsonl.zst
      - train/chunk1/example_train_39.jsonl.zst
      - train/chunk1/example_train_40.jsonl.zst
    split: train
    type: completion
dataset_prepared_path: last_run_mixed
val_set_size: 0.0
output_dir: ./outputs/model-out-mix

sequence_len: 1024
sample_packing: true
# unfrozen_parameters:
#   - model.layers.*.mlp.(gate|up|down)_proj.weight$
wandb_project: whale-v3-post-pt
wandb_entity: theblackcat102
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 32
micro_batch_size: 1
num_epochs: 4
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.00008
max_grad_norm: 10.0
adam_beta2: 0.95

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

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

warmup_steps: 50
evals_per_epoch:
eval_table_size:
save_steps: 100
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens: