Files
sci-gui-agent-benchmark/run_os_symphony.sh
2025-12-23 14:30:44 +08:00

57 lines
2.0 KiB
Bash

EXP_NAME="os-osworld-origin-nogdrive-gpt5-gta1-32b-step50-20251220-ybw"
# enable_rewrite_instruction
python run_multienv_os_symphony.py \
--provider_name "docker" \
--path_to_vm "xxx" \
--headless \
--num_envs 1 \
--max_steps 50 \
--benchmark osworld \
--domain "all" \
--test_all_meta_path evaluation_examples/test_nogdrive.json \
--result_dir "results" \
--region "us-east-1" \
--tool_config mm_agents/os_symphony/tool/all_tool_config.yaml \
--orchestrator_provider "openai" \
--orchestrator_model "gpt-5" \
--orchestrator_url "https://api.boyuerichdata.opensphereai.com/v1" \
--orchestrator_api_key "xxx" \
--orchestrator_temperature 0.1 \
--orchestrator_keep_first_image \
--max_trajectory_length 8 \
--grounder_provider "vllm" \
--grounder_model "gta1_32b" \
--grounder_api_key "none" \
--grounder_url "https://h.pjlab.org.cn/kapi/workspace.kubebrain.io/ailab-intern11/dingzichen-7jzkt-932268-worker-0.dingzichen/18080/v1/" \
--grounding_smart_resize \
--grounding_width 1280 \
--grounding_height 800 \
--coder_provider "openai" \
--coder_model "gpt-5" \
--coder_url "https://api.boyuerichdata.opensphereai.com/v1" \
--coder_api_key "xxx" \
--coder_temperature 0.1 \
--coder_budget 20 \
--memoryer_provider "openai" \
--memoryer_model "gpt-5" \
--memoryer_url "https://api.boyuerichdata.opensphereai.com/v1" \
--memoryer_api_key "xxx" \
--memoryer_temperature 0.1 \
--memoryer_max_images 8 \
--searcher_provider "openai" \
--searcher_model "gpt-5" \
--searcher_url "https://api.boyuerichdata.opensphereai.com/v1" \
--searcher_api_key "xxx" \
--searcher_temperature 0.1 \
--searcher_type "vlm" \
--searcher_engine "duckduckgo" \
--searcher_budget 20\
--searcher_screen_width 1920 \
--searcher_screen_height 1080 \
--searcher_path_to_vm "xxx" \
--sleep_after_execution 3 \
--exp_name ${EXP_NAME} \
--enable_reflection
# bash scripts/remove_all_osworld_container.sh > logs/${EXP_NAME}.log 2>&1 --enable_rewrite_instruction --grounding_smart_resize