| AssemblyWorldBench v1 launch commands |
|
|
| Run from the assembly-world-bench directory after installation and agent authentication. |
| Set AGENT (codex or claude), MODEL and MCP_COMMAND explicitly. Inputs are downloaded |
| from the tool's fixed HF commit; task text and reference conditions cannot be overridden. |
| Each command creates a fresh run directory. Evaluate those independent nonoverlapping |
| block runs together with: |
| uv run assembly-world-bench eval logs/partnet-none/RUN_ID logs/partnet-final-image/RUN_ID logs/ikea-manualbook/RUN_ID logs/assemblybench-manualbook/RUN_ID logs/fantastic-breaks-none/RUN_ID |
|
|
| Alternatively run all five blocks sequentially in one run: |
| uv run assembly-world-bench run --all --agent "$AGENT" --model "$MODEL" --mcp-command "$MCP_COMMAND" --headless |
|
|
| partnet-none: 20 frozen evaluations |
| uv run assembly-world-bench run --block partnet-none --agent "$AGENT" --model "$MODEL" --mcp-command "$MCP_COMMAND" --headless --logs logs/partnet-none |
|
|
| partnet-final-image: 20 frozen evaluations |
| uv run assembly-world-bench run --block partnet-final-image --agent "$AGENT" --model "$MODEL" --mcp-command "$MCP_COMMAND" --headless --logs logs/partnet-final-image |
|
|
| ikea-manualbook: 20 frozen evaluations |
| uv run assembly-world-bench run --block ikea-manualbook --agent "$AGENT" --model "$MODEL" --mcp-command "$MCP_COMMAND" --headless --logs logs/ikea-manualbook |
|
|
| assemblybench-manualbook: 20 frozen evaluations |
| uv run assembly-world-bench run --block assemblybench-manualbook --agent "$AGENT" --model "$MODEL" --mcp-command "$MCP_COMMAND" --headless --logs logs/assemblybench-manualbook |
|
|
| fantastic-breaks-none: 20 frozen evaluations |
| uv run assembly-world-bench run --block fantastic-breaks-none --agent "$AGENT" --model "$MODEL" --mcp-command "$MCP_COMMAND" --headless --logs logs/fantastic-breaks-none |
|
|
|
|
