#!/bin/bash BEAKER_IMAGE=${1:-nathanl/open_instruct_auto} # our released olmo3 thinker model actually doesnt use pipelinerl, but using it should do just as well. python mason.py \ --cluster ai2/jupiter \ --image hamishivi/open_instruct_testing_1110 \ --pure_docker_mode \ --workspace ai2/olmo-instruct \ --priority urgent \ --preemptible \ --num_nodes 9 \ --gpus 8 \ --max_retries 0 \ --env RAY_CGRAPH_get_timeout=300 \ --env VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 \ --env HOSTED_VLLM_API_BASE=http://ceres-cs-aus-447.reviz.ai2.in:8001/v1 \ -- source configs/beaker_configs/ray_node_setup.sh \&\& source configs/beaker_configs/code_api_setup.sh \&\& python open_instruct/grpo_fast.py \ --exp_name 7b_olmo3_thinker_no_pipeline \ --beta 0.0 \ --num_samples_per_prompt_rollout 8 \ --num_unique_prompts_rollout 64 \ --num_mini_batches 1 \ --num_epochs 1 \ --learning_rate 1e-6 \ --per_device_train_batch_size 1 \ --output_dir /output \ --kl_estimator 2 \ --dataset_mixer_list allenai/Dolci-Think-RL-7B 1.0 \ --dataset_mixer_list_splits train \ --dataset_mixer_eval_list allenai/Dolci-Think-RL-7B 8 \ --dataset_mixer_eval_list_splits train \ --max_token_length 10240 \ --max_prompt_token_length 2048 \ --response_length 32768 \ --pack_length 35840 \ --model_name_or_path allenai/Olmo-3-7B-Think-DPO \ --chat_template_name olmo_thinker \ --non_stop_penalty False \ --mask_truncated_completions False \ --temperature 1.0 \ --ground_truths_key ground_truth \ --sft_messages_key messages \ --total_episodes 10000000 \ --deepspeed_stage 3 \ --num_learners_per_node 8 8 \ --vllm_num_engines 56 \ --vllm_tensor_parallel_size 1 \ --lr_scheduler_type constant \ --apply_verifiable_reward true \ --seed 1 \ --local_eval_every 50 \ --save_freq 25 \ --beaker_eval_freq 50 \ --eval_priority urgent \ --try_launch_beaker_eval_jobs_on_weka True \ --gradient_checkpointing \ --with_tracking \ --llm_judge_model hosted_vllm/Qwen/Qwen3-32B \ --llm_judge_timeout 600 \ --llm_judge_max_tokens 2048 \ --llm_judge_max_context_length 32768 \ --code_api_url https://p9f1719l7f.execute-api.us-west-2.amazonaws.com/prod/test_program \ --code_pass_rate_reward_threshold 0.99 \ --oe_eval_max_length 32768 \ --checkpoint_state_freq 100 \ --backend_timeout 1200 \ --inflight_updates false \ --async_steps 1 \ --oe_eval_beaker_image oe-eval-beaker/oe_eval_olmo2_retrofit_auto \ --oe_eval_tasks mmlu:cot::hamish_zs_reasoning_deepseek,popqa::hamish_zs_reasoning_deepseek,simpleqa::tulu-thinker_deepseek,bbh:cot::hamish_zs_reasoning_deepseek_v2,gpqa:0shot_cot::hamish_zs_reasoning_deepseek,zebralogic::hamish_zs_reasoning_deepseek,agi_eval_english:0shot_cot::hamish_zs_reasoning_deepseek,minerva_math::hamish_zs_reasoning_deepseek,minerva_math_500::hamish_zs_reasoning_deepseek,gsm8k::zs_cot_latex_deepseek,omega_500:0-shot-chat_deepseek,aime:zs_cot_r1::pass_at_32_2024_deepseek,aime:zs_cot_r1::pass_at_32_2025_deepseek,codex_humanevalplus:0-shot-chat::tulu-thinker_deepseek,mbppplus:0-shot-chat::tulu-thinker_deepseek,livecodebench_codegeneration::tulu-thinker_deepseek,alpaca_eval_v3::hamish_zs_reasoning_deepseek,ifeval::hamish_zs_reasoning_deepseek