| set -x | |
| export VLLM_ATTENTION_BACKEND=FLASH_ATTN | |
| export RAY_memory_monitor_refresh_ms=0 | |
| export RAY_LOGGING_LEVEL=DEBUG | |
| export HYDRA_FULL_ERROR=1 | |
| export PYTHONPATH="${PYTHONPATH}:$(pwd)/verl" | |
| OUTPUT_SEED_PATH=${OUTPUT_SEED_PATH:-data/3b_coder_seed_io.jsonl} | |
| OUTPUT_ERROR_SEED_PATH=${OUTPUT_ERROR_SEED_PATH:-data/3b_coder_error_seed_io.jsonl} | |
| OUTPUT_CODE_F_SEED_PATH=${OUTPUT_CODE_F_SEED_PATH:-data/3b_coder_code_f_seed_io.jsonl} | |
| python -m absolute_zero_reasoner.main_azr_ppo \ | |
| data.shuffle=True \ | |
| actor_rollout_ref.ref.include_ref=False \ | |
| algorithm.adv_estimator=reinforce_plus_plus \ | |
| data.train_files=data/code_reason/test_answer.parquet \ | |
| data.val_files=data/code_reason/test_answer.parquet \ | |
| data.train_batch_size=64 \ | |
| data.val_batch_size=1312 \ | |
| data.max_prompt_length=6144 \ | |
| data.max_response_length=8096 \ | |
| azr.data_selection_strategy.content_max_length=5600 \ | |
| actor_rollout_ref.model.path=Qwen/Qwen2.5-Coder-3B \ | |
| actor_rollout_ref.actor.optim.lr=1e-6 \ | |
| actor_rollout_ref.model.use_remove_padding=True \ | |
| actor_rollout_ref.actor.ppo_mini_batch_size=128 \ | |
| actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=16 \ | |
| actor_rollout_ref.actor.use_kl_loss=False \ | |
| actor_rollout_ref.actor.kl_loss_coef=0.0 \ | |
| actor_rollout_ref.actor.kl_loss_type=low_var_kl \ | |
| actor_rollout_ref.actor.ulysses_sequence_parallel_size=2 \ | |
| actor_rollout_ref.model.enable_gradient_checkpointing=True \ | |
| actor_rollout_ref.model.pretrained_tokenizer=True \ | |
| actor_rollout_ref.actor.fsdp_config.param_offload=False \ | |
| actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \ | |
| actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=64 \ | |
| actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=64 \ | |
| actor_rollout_ref.rollout.tensor_model_parallel_size=2 \ | |
| actor_rollout_ref.rollout.name=vllm \ | |
| actor_rollout_ref.rollout.max_num_batched_tokens=16384 \ | |
| actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \ | |
| actor_rollout_ref.rollout.enforce_eager=False \ | |
| actor_rollout_ref.rollout.free_cache_engine=False \ | |
| actor_rollout_ref.rollout.n=1 \ | |
| actor_rollout_ref.rollout.temperature=1.0 \ | |
| actor_rollout_ref.ref.fsdp_config.param_offload=True \ | |
| algorithm.kl_ctrl.kl_coef=0.0 \ | |
| trainer.critic_warmup=0 \ | |
| trainer.logger=['console'] \ | |
| trainer.project_name='azr' \ | |
| trainer.experiment_name='test' \ | |
| trainer.n_gpus_per_node=2 \ | |
| trainer.nnodes=1 \ | |
| trainer.save_freq=10 \ | |
| trainer.remove_previous_ckpt_in_save=True \ | |
| trainer.del_local_ckpt_after_load=True \ | |
| trainer.test_freq=10 \ | |
| +trainer.val_before_train=False \ | |
| reward_fn.extraction_type=answer_conditional \ | |
| reward_fn.math_metric=math_verify \ | |
| trainer.log_val_generations=0 \ | |
| azr.data_selection_strategy.update_iteration=1 \ | |
| azr.seed_dataset=data/3b_coder_seed_io.jsonl \ | |
| azr.output_seed_path=data/3b_coder_seed_io.jsonl \ | |
| azr.error_seed_dataset=data/3b_coder_error_seed_io.jsonl \ | |
| azr.output_error_seed_path=data/3b_coder_error_seed_io.jsonl \ | |
| azr.code_f_seed_dataset=data/3b_coder_code_f_seed_io.jsonl \ | |
| azr.output_code_f_seed_path=data/3b_coder_code_f_seed_io.jsonl \ | |
| azr.pretrain_pred_steps=-1 \ | |
| azr.executor=qwq \ | |
| azr.ast_check=True \ | |
| azr.reward.n_samples=8 \ | |
| azr.problem_types=['code_i','code_o','code_f'] \ | |
| azr.data_selection_strategy.banned_keywords_for_errors_and_exceptions=['raise'] \ | |
| trainer.debug=True \ | |
| trainer.debug_port=6380 \ | |
| azr.reward.generation_reward_config.complexity_reward.coef=0.0 \ | |
| azr.reward.generation_reward_config.complexity_reward.max=0.0 \ | |
| azr.reward.generation_reward_config.complexity_reward.enabled=False \ | |
| azr.reward.generation_reward_config.mean_edit_distance_reward.coef=0.0 \ | |
| azr.reward.generation_reward_config.mean_edit_distance_reward.max=0.0 \ | |
| azr.reward.generation_reward_config.mean_edit_distance_reward.enabled=False \ | |
| azr.reward.generation_reward_config.halstead_reward.coef=0.0 \ | |
| azr.reward.generation_reward_config.halstead_reward.max=0.0 \ | |
| azr.reward.generation_reward_config.halstead_reward.enabled=False \ | |
| azr.reward.generation_reward_config.answer_diversity_reward.coef=0.0 \ | |
| azr.reward.generation_reward_config.answer_diversity_reward.max=0.0 \ | |
| azr.reward.generation_reward_config.answer_diversity_reward.enabled=False \ | |
| azr.reward.generation_reward_config.answer_diversity_reward.hierarchical=False \ | |
| azr.pred_data_mix_strategy=max_new \ | |
| azr.data_selection_strategy.seed_batch_factor=4 \ | |
| azr.data_selection_strategy.valid_program_filter=all \ | |
| azr.data_selection_strategy.max_programs=16384 \ | |
| azr.data_selection_strategy.batched_estimate=False \ | |
| azr.reward.generation_reward_config.intrinsic_combine_method=sum \ | |
| azr.gen_data_probabilities_strategy=uniform \ | |
| trainer.resume_mode=auto \ | |
| azr.data_selection_strategy.composite_start_step=-1 \ | |
| azr.data_selection_strategy.composite_chance=0.0 \ | |
| azr.reward.generation_reward_config.remove_comments=False \ | |
| azr.reward.generation_reward_config.remove_after_return=False \ | |
| azr.reward.generation_reward_config.use_original_code_as_ref=True \ | |
| azr.reward.generation_reward_config.remove_print=False \ | |
| azr.data_selection_strategy.composite_function_n_min=0 \ | |
| azr.data_selection_strategy.composite_function_n_max=0 \ | |
| azr.reward.code_f_reward_type=binary \ | |
| trainer.wandb_run_id=null \ | |
| trainer.debug=True \ | |
| trainer.debug_port=6380 \ | |
| trainer.total_epochs=30 $@ |