base_model: IntervitensInc/gemma-2-9b-chatml
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: true
strict: false

# huggingface repo
datasets:
  - path: NewEden/Kalo-Opus-Instruct-25K-Refusal-killed
    type: chatml
  - path: NewEden/Gryphe-3.5-16k-Subset
    type: chatml

val_set_size: 0.05
output_dir: ./outputs/out

adapter: lora
peft_use_rslora: true
lora_r: 32
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true

sequence_len: 8192
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

wandb_project: magnum 9b
wandb_entity:
wandb_watch:
wandb_name: magnum 9b inst 
wandb_log_model:


gradient_accumulation_steps: 32
micro_batch_size: 1
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.00002

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

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

warmup_ratio: 0.1
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 1
debug:
deepspeed: zero2.json
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens: