# check ./tina/utils/constant.py
model_post_train_dataset_name: curated_rg_math
model_post_train_type: grpo
rl_post_train_reward_funcs:
- format
- accuracy
rl_post_train_reward_weights:
- 1.0
- 2.0


# Model configs from trl
model_name_or_path: DeepSeek-R1-Distill-Qwen-1.5B
attn_implementation: flash_attention_2
use_peft: true
lora_r: 32
lora_alpha: 128
lora_dropout: 0.05
lora_target_modules:
- q_proj
- k_proj
- v_proj
- o_proj
- down_proj
- up_proj
- gate_proj


# GRPO trainer configs from trl
bf16: true
use_vllm: true
vllm_device: cuda:0
vllm_gpu_memory_utilization: 0.4
vllm_max_model_len: 4608
gradient_accumulation_steps: 4
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
log_level: info
logging_first_step: true
logging_steps: 1
logging_strategy: steps
push_to_hub: true
hub_strategy: every_save
hub_private_repo: true
hub_model_id: starzmustdie

learning_rate: 1e-06
lr_scheduler_type: cosine_with_min_lr
lr_scheduler_kwargs:
  min_lr_rate: 0.1
max_prompt_length: 512
max_completion_length: 3584
max_steps: 2500 # use 2500 for lr scheduler but stop at 1250 steps
num_generations: 4
num_train_epochs: 1
overwrite_output_dir: true
per_device_train_batch_size: 4
report_to:
- wandb

save_strategy: steps
save_steps: 100
save_total_limit: 100

seed: 42
temperature: 0.7
warmup_ratio: 0.1