recipes/Qwen2.5-Coder-7B-Instruct/grpo/config_codeforces.yaml

# Model arguments
model_name_or_path: Qwen/Qwen2.5-Coder-7B-Instruct
model_revision: main
torch_dtype: bfloat16
attn_implementation: flash_attention_2
# Data training arguments
dataset_name: open-r1/codeforces
dataset_prompt_column: prompt
dataset_config: verifiable-prompts
dataset_test_split: test
dataset_train_split: train

system_prompt: "You are a helpful AI Assistant that provides well-reasoned and detailed responses. You first think about the reasoning process as an internal monologue and then provide the user with the answer. Respond in the following format: <think>\n...\n</think>\n<answer>\n...\n</answer>"

# GRPO trainer config
callbacks:
- push_to_hub_revision
benchmarks:
- lcb_v4
beta: 0.0
loss_type: dr_grpo
scale_rewards: false
bf16: true
do_eval: false
eval_strategy: "no"
use_vllm: true
vllm_device: auto
vllm_gpu_memory_utilization: 0.7
gradient_accumulation_steps: 32
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
hub_model_id: open-r1/Qwen2.5-Coder-7B-Instruct-Codeforces-GRPO
hub_model_revision: v01.00
hub_strategy: every_save
learning_rate: 1.0e-06
log_completions: true
log_level: info
logging_first_step: true
logging_steps: 1
logging_strategy: steps
lr_scheduler_type: constant_with_warmup
max_grad_norm: 0.2
max_prompt_length: 2000
max_completion_length: 8192
max_steps: -1
num_generations: 16
# aiming for 1k optimization steps
# total_samples_per_batch = num_gpus * grad_accumulation_steps * per_device_batch_size = 8 * 32 * 4 = 1024
# unique_prompts_per_batch = total_samples_per_batch / num_generations = 1024 / 16 = 64
# #dataset ~= 16k (8k * 2, for python and cpp)
# global_steps_per_epoch = #dataset / unique_prompts_per_batch = 16k / 64 ~= 250
# epochs_for_1k_steps = 1000/250 = 4 epochs
num_train_epochs: 4
output_dir: data/Qwen2.5-Coder-7B-Instruct-Codeforces-GRPO_v01.00
overwrite_output_dir: true
per_device_train_batch_size: 4
push_to_hub: true
report_to:
- wandb
reward_funcs:
- cf_code
- code_format
reward_weights:
- 1.0
- 0.1
save_strategy: "steps"
save_steps: 0.05
save_total_limit: 1
seed: 42
temperature: 0.7
wandb_entity: huggingface
wandb_project: open-r1
warmup_ratio: 0.1

mask_truncated_completions: true
# for each generation, evaluate these many test cases in parallel, then check if any of them failed (0 score): if so stop evaluating
# otherwise continue with the next batch of test cases. Useful to avoid overloading the eval server + save time on wrong solutions
code_eval_test_batch_size: -1
code_eval_scoring_mode: weighted_sum