CPT training is giving pretty unstalbe results with the learning rate 2e-5.

[shamanez]

I am trying to conduct CPT with a mistral-instruct-v2. But every time, I notice an overshooting in the grad norm. I tried different datasets and managed to re-produce the same issue.

I am using 8 80GB GPUs, and my effective batch size is 1024.

My config:

# Model arguments
model_name_or_path: mistralai/Mistral-7B-Instruct-v0.2
model_revision: main
torch_dtype: bfloat16

# Data training arguments
dataset_mixer:
  arcee-ai/sec-data-full: 1.0
dataset_splits:
  - train
preprocessing_num_workers: 12
text_column: "text"

# SFT trainer config
bf16: true
do_eval: False
evaluation_strategy: "no"
gradient_accumulation_steps: 64
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: False
hub_model_id: arcee-ai/mistral-instruct-v2-sec
hub_strategy: every_save
learning_rate: 2.0e-04
log_level: info
logging_steps: 1  
logging_strategy: steps
lr_scheduler_type: cosine
max_seq_length: 4096
max_steps: -1
num_train_epochs: 1
output_dir: data/mistral-instruct-v2-sec-expanded
overwrite_output_dir: true
per_device_eval_batch_size: 1
per_device_train_batch_size: 4
push_to_hub: true 
remove_unused_columns: true
report_to:
- wandb
save_strategy: 'steps'
save_steps: 50
save_total_limit: 2
seed: 42
warmup_ratio: 0.1

[shamanez]