cot_sft.yaml

### model
model_name_or_path: CodeLlama-7B-evol

### method
stage: sft
do_train: true
finetuning_type: full
flash_attn: auto
deepspeed: examples/deepspeed/ds_z3_config.json

### dataset
dataset: sft_COT_safecoder_v4,sft_COT_safecoder_v4_autocomplete
# new_special_tokens: <think>,</think>
template: default
cutoff_len: 4096
max_samples: 1000000
overwrite_cache: true
preprocessing_num_workers: 16
seed: 42

### output

output_dir: saves/CodeLlama-7B/full/CodeLlama-7B-safetycot-2025-03-14-15-00
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
report_to: tensorboard

### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 4
learning_rate: 1.0e-5
num_train_epochs: 3.0
optim: adamw_torch
lr_scheduler_type: cosine
warmup_ratio: 0.1
# warmup_steps: 100
bf16: true
max_grad_norm: 1.0
ddp_timeout: 180000000

### eval
do_eval: false
# val_size: 0.1
# per_device_eval_batch_size: 1
# eval_strategy: steps
# eval_steps: 100

include_num_input_tokens_seen: true