recipes/OlympicCoder-32B/sft/config_v00.00.yaml

# Config for 16 nodes of 8 H100s with FSDP1
# Model arguments
model_name_or_path: Qwen/Qwen2.5-Coder-32B-Instruct
model_revision: main
torch_dtype: bfloat16
attn_implementation: flash_attention_2

# Data training arguments
dataset_name: open-r1/codeforces-cots
dataset_config: solutions_decontaminated
dataset_num_proc: 12

# SFT trainer config
bf16: true
do_eval: false
eval_strategy: 'no'
gradient_accumulation_steps: 1
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
hub_always_push: true
hub_model_id: OlympicCoder-32B
hub_strategy: every_save
learning_rate: 4.0e-05
log_level: info
logging_steps: 1
logging_strategy: steps
lr_scheduler_type: cosine_with_min_lr
lr_scheduler_kwargs:
  min_lr_rate: 0.1
packing: false
max_grad_norm: 0.2
max_length: 22528 # we were unable to train at 32k due to OOM. See https://github.com/huggingface/transformers/issues/35983 for context parallelism support.
max_steps: -1
num_train_epochs: 10
optim: paged_adamw_8bit
output_dir: data/OlympicCoder-32B
overwrite_output_dir: true
per_device_eval_batch_size: 1
per_device_train_batch_size: 1
push_to_hub: true
report_to:
- wandb
save_only_model: true # needed to bypass FSDP errors with saving paged optimizers
save_strategy: epoch
save_total_limit: 1
seed: 42
use_liger_kernel: false # fails on multi-node
warmup_ratio: 0.03