# vllm serve meta-llama/Llama-3.3-70B-Instruct --port 8004 --tensor-parallel-size 4 --max-model-len 16384 --gpu-memory-utilization 0.9 # python -m sglang.launch_server --model-path meta-llama/Llama-3.3-70B-Instruct --port 8004 --tp-size 4 --context-length 16384 # python -m sglang.launch_server --model-path meta-llama/Llama-3.1-8B-Instruct --port 8003 --tp-size 4 --context-length 16384 # python -m sglang.launch_server --model-path Qwen/Qwen2.5-7B-Instruct --port 8003 --tp-size 4 --context-length 16384 # SFT Models # python -m sglang.launch_server --model-path /shared/storage-01/users/mehri2/LLaMA-Factory/saves/llama-3.1-8b-instruct/full/sft_session_level_reflection/checkpoint-628 --served-model-name meta-llama/Llama-3.1-8B-Instruct --port 8003 --tp-size 4 --context-length 16384 # python -m sglang.launch_server --model-path /shared/storage-01/users/mehri2/LLaMA-Factory/saves/qwen2.5-7b/full/sft_session_level_reflection/checkpoint-628 --served-model-name Qwen/Qwen2.5-7B-Instruct --port 8003 --tp-size 4 --context-length 16384 # GRPO Models # python -m verl.model_merger merge \ # --backend fsdp \ # --local_dir /shared/storage-01/users/mehri2/mem/collaborativeagents/training/grpo_verl/results/v3/global_step_200/actor \ # --target_dir /shared/storage-01/users/mehri2/mem/collaborativeagents/training/grpo_verl/results/v3/global_step_200_merged_hf # python -m sglang.launch_server --model-path /shared/storage-01/users/mehri2/mem/collaborativeagents/training/grpo_verl/results/v3/global_step_200_merged_hf --served-model-name meta-llama/Llama-3.1-8B-Instruct --port 8003 --tp-size 4 --context-length 16384 BATCH_SIZE=100 BATCH_SIZE=50 # Loop over eval sizes and datasets for EVAL_SIZE in 20; do for DATASET in math-hard math-500 logiqa mmlu medqa; do # humaneval bigcodebench # Convert dataset name for file paths (replace - with _) DATASET_FILE=$(echo ${DATASET} | tr '-' '_') echo "Running experiments for dataset: ${DATASET} with eval_size ${EVAL_SIZE}" # # no_user experiment # python3 run.py --experiment_type no_user --dataset ${DATASET} --eval_size ${EVAL_SIZE} --batch_size ${BATCH_SIZE} \ # --collaborator_model_name hosted_vllm/meta-llama/Llama-3.1-8B-Instruct --collaborator_api_base http://localhost:8003/v1 --collaborator_api_key EMPTY \ # --judge_model_name hosted_vllm/meta-llama/Llama-3.3-70B-Instruct --judge_api_base http://localhost:8004/v1 --judge_api_key EMPTY \ # --output_file /shared/storage-01/users/mehri2/mem/collaborativeagents/scripts/runs/llama70b_temp_1_llama8b/no_user/${DATASET_FILE}_llama70b_user_llama8b_agent_no_user_eval_size_${EVAL_SIZE}.jsonl \ # >> ./runs/llama70b_temp_1_llama8b/no_user/${DATASET_FILE}_llama70b_user_llama8b_agent_no_user_eval_size_${EVAL_SIZE}.out 2>&1 # # user_no_profile experiment # python3 run.py --experiment_type user_no_profile --dataset ${DATASET} --eval_size ${EVAL_SIZE} --max_turns 10 --batch_size ${BATCH_SIZE} \ # --user_model_name hosted_vllm/meta-llama/Llama-3.3-70B-Instruct --user_api_base http://localhost:8004/v1 --user_api_key EMPTY \ # --collaborator_model_name hosted_vllm/meta-llama/Llama-3.1-8B-Instruct --collaborator_api_base http://localhost:8003/v1 --collaborator_api_key EMPTY \ # --judge_model_name hosted_vllm/meta-llama/Llama-3.3-70B-Instruct --judge_api_base http://localhost:8004/v1 --judge_api_key EMPTY \ # --output_file /shared/storage-01/users/mehri2/mem/collaborativeagents/scripts/runs/llama70b_temp_1_llama8b/user_no_profile/${DATASET_FILE}_llama70b_user_llama8b_agent_user_no_profile_eval_size_${EVAL_SIZE}.jsonl \ # >> ./runs/llama70b_temp_1_llama8b/user_no_profile/${DATASET_FILE}_llama70b_user_llama8b_agent_user_no_profile_eval_size_${EVAL_SIZE}.out 2>&1 # # user_profiles_without_preferences experiment # python3 run.py --experiment_type user_profiles_without_preferences --dataset ${DATASET} --eval_size ${EVAL_SIZE} --max_turns 10 --batch_size ${BATCH_SIZE} \ # --user_model_name hosted_vllm/meta-llama/Llama-3.3-70B-Instruct --user_api_base http://localhost:8004/v1 --user_api_key EMPTY \ # --collaborator_model_name hosted_vllm/meta-llama/Llama-3.1-8B-Instruct --collaborator_api_base http://localhost:8003/v1 --collaborator_api_key EMPTY \ # --judge_model_name hosted_vllm/meta-llama/Llama-3.3-70B-Instruct --judge_api_base http://localhost:8004/v1 --judge_api_key EMPTY \ # --output_file /shared/storage-01/users/mehri2/mem/collaborativeagents/scripts/runs/llama70b_temp_1_llama8b/user_profiles_without_preferences/${DATASET_FILE}_llama70b_user_llama8b_agent_user_profiles_without_preferences_eval_size_${EVAL_SIZE}.jsonl \ # >> ./runs/llama70b_temp_1_llama8b/user_profiles_without_preferences/${DATASET_FILE}_llama70b_user_llama8b_agent_user_profiles_without_preferences_eval_size_${EVAL_SIZE}.out 2>&1 # user_profiles_with_preferences experiment python3 run.py --experiment_type user_profiles_with_preferences --dataset ${DATASET} --eval_size ${EVAL_SIZE} --max_turns 10 --batch_size ${BATCH_SIZE} \ --user_model_name hosted_vllm/meta-llama/Llama-3.3-70B-Instruct --user_api_base http://localhost:8004/v1 --user_api_key EMPTY \ --collaborator_model_name hosted_vllm/meta-llama/Llama-3.1-8B-Instruct --collaborator_api_base http://localhost:8003/v1 --collaborator_api_key EMPTY \ --judge_model_name hosted_vllm/meta-llama/Llama-3.3-70B-Instruct --judge_api_base http://localhost:8004/v1 --judge_api_key EMPTY \ --output_file /shared/storage-01/users/mehri2/mem/collaborativeagents/scripts/runs/llama70b_temp_1_llama8b_grpo_v3_ckpt200/user_profiles_with_preferences/${DATASET_FILE}_llama70b_user_llama8b_agent_user_profiles_with_preferences_eval_size_${EVAL_SIZE}.jsonl \ >> /shared/storage-01/users/mehri2/mem/collaborativeagents/scripts/runs/llama70b_temp_1_llama8b_grpo_v3_ckpt200/user_profiles_with_preferences/${DATASET_FILE}_llama70b_user_llama8b_agent_user_profiles_with_preferences_eval_size_${EVAL_SIZE}.out 2>&1 # # agent_with_user_preferences experiment # python3 run.py --experiment_type agent_with_user_preferences --dataset ${DATASET} --eval_size ${EVAL_SIZE} --max_turns 10 --batch_size ${BATCH_SIZE} \ # --user_model_name hosted_vllm/meta-llama/Llama-3.3-70B-Instruct --user_api_base http://localhost:8004/v1 --user_api_key EMPTY \ # --collaborator_model_name hosted_vllm/meta-llama/Llama-3.1-8B-Instruct --collaborator_api_base http://localhost:8003/v1 --collaborator_api_key EMPTY \ # --judge_model_name hosted_vllm/meta-llama/Llama-3.3-70B-Instruct --judge_api_base http://localhost:8004/v1 --judge_api_key EMPTY \ # --output_file /shared/storage-01/users/mehri2/mem/collaborativeagents/scripts/runs/llama70b_temp_1_llama8b/agent_with_user_preferences/${DATASET_FILE}_llama70b_user_llama8b_agent_agent_with_user_preferences_eval_size_${EVAL_SIZE}_v2.jsonl \ # >> ./runs/llama70b_temp_1_llama8b/agent_with_user_preferences/${DATASET_FILE}_llama70b_user_llama8b_agent_agent_with_user_preferences_eval_size_${EVAL_SIZE}_v2.out 2>&1 # # agent_with_reflection experiment # python3 run.py --experiment_type agent_with_reflection --dataset ${DATASET} --eval_size ${EVAL_SIZE} --max_turns 10 --batch_size ${BATCH_SIZE} \ # --user_model_name hosted_vllm/meta-llama/Llama-3.3-70B-Instruct --user_api_base http://localhost:8004/v1 --user_api_key EMPTY \ # --collaborator_model_name hosted_vllm/meta-llama/Llama-3.1-8B-Instruct --collaborator_api_base http://localhost:8003/v1 --collaborator_api_key EMPTY \ # --judge_model_name hosted_vllm/meta-llama/Llama-3.3-70B-Instruct --judge_api_base http://localhost:8004/v1 --judge_api_key EMPTY \ # --output_file /shared/storage-01/users/mehri2/mem/collaborativeagents/scripts/runs/llama70b_temp_1_llama8b/agent_with_reflection_v3/${DATASET_FILE}_llama70b_user_llama8b_agent_agent_with_reflection_eval_size_${EVAL_SIZE}.jsonl \ # >> ./runs/llama70b_temp_1_llama8b/agent_with_reflection_v3/${DATASET_FILE}_llama70b_user_llama8b_agent_agent_with_reflection_eval_size_${EVAL_SIZE}.out 2>&1 # # agent_with_reflection_and_scaffolding # python3 run.py --experiment_type agent_with_reflection_and_scaffolding --dataset ${DATASET} --eval_size ${EVAL_SIZE} --max_turns 10 --batch_size ${BATCH_SIZE} \ # --user_model_name hosted_vllm/meta-llama/Llama-3.3-70B-Instruct --user_api_base http://localhost:8004/v1 --user_api_key EMPTY \ # --collaborator_model_name hosted_vllm/meta-llama/Llama-3.1-8B-Instruct --collaborator_api_base http://localhost:8003/v1 --collaborator_api_key EMPTY \ # --judge_model_name hosted_vllm/meta-llama/Llama-3.3-70B-Instruct --judge_api_base http://localhost:8004/v1 --judge_api_key EMPTY \ # --output_file /shared/storage-01/users/mehri2/mem/collaborativeagents/scripts/runs/llama70b_temp_1_llama8b/agent_with_reflection_and_scaffolding/${DATASET_FILE}_llama70b_user_llama8b_agent_agent_with_reflection_and_scaffolding_eval_size_${EVAL_SIZE}.jsonl \ # >> ./runs/llama70b_temp_1_llama8b/agent_with_reflection_and_scaffolding/${DATASET_FILE}_llama70b_user_llama8b_agent_agent_with_reflection_and_scaffolding_eval_size_${EVAL_SIZE}.out 2>&1 # agent_with_reflection_and_proper_scaffolding python3 run.py --experiment_type agent_with_reflection_and_proper_scaffolding --dataset ${DATASET} --eval_size ${EVAL_SIZE} --max_turns 10 --batch_size ${BATCH_SIZE} \ --user_model_name hosted_vllm/meta-llama/Llama-3.3-70B-Instruct --user_api_base http://localhost:8004/v1 --user_api_key EMPTY \ --collaborator_model_name hosted_vllm/meta-llama/Llama-3.1-8B-Instruct --collaborator_api_base http://localhost:8003/v1 --collaborator_api_key EMPTY \ --judge_model_name hosted_vllm/meta-llama/Llama-3.3-70B-Instruct --judge_api_base http://localhost:8004/v1 --judge_api_key EMPTY \ --output_file /shared/storage-01/users/mehri2/mem/collaborativeagents/scripts/runs/llama70b_temp_1_llama8b_grpo_v3_ckpt200/agent_with_reflection_and_proper_scaffolding/${DATASET_FILE}_llama70b_user_llama8b_agent_agent_with_reflection_and_proper_scaffolding_eval_size_${EVAL_SIZE}.jsonl \ >> /shared/storage-01/users/mehri2/mem/collaborativeagents/scripts/runs/llama70b_temp_1_llama8b_grpo_v3_ckpt200/agent_with_reflection_and_proper_scaffolding/${DATASET_FILE}_llama70b_user_llama8b_agent_agent_with_reflection_and_proper_scaffolding_eval_size_${EVAL_SIZE}.out 2>&1 done done