| [ |
| { |
| "name": "inspect_evals/zerobench - Sample 86", |
| "description": "Task from inspect_evals/zerobench dataset, sample 86", |
| "prompt": "Count the number of wall tiles based on the following rules:\n\n- Only count wall tiles, not floor tiles\n- A partial tile of of any amount counts as 1\n- Only count tiles that are not split horizontally\n- If something is in front of the tile and part of it is still visible, still count that tile\n- Behind the toilet is tiled, and we need to count the tiles obscured by it.\n\nHow many tiles are there based on the above tiles?\n\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench", |
| "sample_id": "86", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench - Sample 74", |
| "description": "Task from inspect_evals/zerobench dataset, sample 74", |
| "prompt": "(1) In which year did the all-time top goal scorer associated with the crest on the right lose an important match to the national team whose primary spoken language is the one written on the left object? \n(2) As of January 2025, how many fewer Ballon D\u2019Ors does the person who won the Ballon D\u2019or in the year of the previous answer, have compared to the person associated with the crest on the right? \n(3) In which minute of the associated match from (1) did a player score the winning goal? \n(4) How many joules of energy would be stored in 1kg of the food item on the left?\n\nSum the answers to 1, 2, 3 and 4 to give the final answer.\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench", |
| "sample_id": "74", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench - Sample 9", |
| "description": "Task from inspect_evals/zerobench dataset, sample 9", |
| "prompt": "The picture contains three columns of mirrored words.\n\nCompute the product of the squares of (i) the number of 4 legged animals in column 1, (ii) the number of South American cities in column 2 and (iii) the number of elements with atomic numbers less than Neon in column 3.\n\nMultiply this number by the number of rows where all items in the row match a condition.\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench", |
| "sample_id": "9", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench - Sample 24", |
| "description": "Task from inspect_evals/zerobench dataset, sample 24", |
| "prompt": "What is the sequence of moves to achieve the highest score? \nBased on the current state of the classic Tetris game, and based on the following assumptions:\n\u2022\tThe only pieces that can be used are the current piece and the next piece (displayed on the right)\n\u2022\tA shape will keep moving downwards by 1 square every second\n\u2022\tThe left key moves the piece 1 square to the left\n\u2022\tRight key moves the piece 1 square to the right\n\u2022\tUp rotates the piece 90 degrees clockwise, and maintains the alignment to the right\n\u2022\tA shape will stop moving when it has become in contact with an existing shape or the bottom of the grid. This will trigger the next shape to start falling\n\u2022\tThe next shape will drop starting from the far left at the top, with its displayed orientation\n\u2022\tThe score is listed in the top right box, currently displayed as 0015400\n\u2022\tClearing a line requires all the boxes in a horizontal line to be filled with shapes and will delete that line of boxes\n\u2022\tClearing a single line with one piece will grant 100 points\n\u2022\tClearing two lines with one piece will grant 300 points\n\u2022\tClearing three lines with one piece will grant 500 points\n\u2022\tClearing four lines (Tetris!) will grant 800 points\n\nGive the sequence of moves with any lateral movement first, followed by any rotations for each piece.\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench", |
| "sample_id": "24", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench - Sample 8", |
| "description": "Task from inspect_evals/zerobench dataset, sample 8", |
| "prompt": "Consider the following:\n(1) How many figurines do not have hair on their head?\n(2) Considering just the number of studs that are visible (not including rows and columns that are completely covered), multiply the length by the width of the gray tile floor.\n(3) What is the maximum number of figures that have the same styled outfit (where the colour scheme can differ but the design and style are the same)?\n(4) How many figurines are there in total?\nCompute the product of the answers to (1), (2), (3) and (4). Note, include the figurine in the background in your calculations.\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench", |
| "sample_id": "8", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench - Sample 88", |
| "description": "Task from inspect_evals/zerobench dataset, sample 88", |
| "prompt": "Nonogram rules: \nColor nonograms are logic puzzles played on a grid, where the goal is to fill squares with the correct colors to reveal a hidden image. Each row and column has sequences of numbers paired with colors, indicating the length and color of consecutive filled squares (called blocks). Blocks of the same color must be separated by at least one empty square, while blocks of different colors can touch. Players use these clues to deduce the placement and color of each square, ensuring all rows and columns match their clues. The puzzle is complete when the entire grid forms the intended multicolored picture. What picture will be depicted in the above image after completing the puzzle? Just provide the name of the object/thing depicted in the image.\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench", |
| "sample_id": "88", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench - Sample 19", |
| "description": "Task from inspect_evals/zerobench dataset, sample 19", |
| "prompt": "Different words are written on the paper in a variety of directions.\nOnly words that are sports score points. Each letter in the words is worth 1 point. Racket sports are worth an additional 3 points. If the word is to the left of a word that is exactly 4 letters in length, it scores an extra 2 points. The final value of each word is calculated as the total number of points multiplied by the number of vertical lines the word crosses.\nCalculate the total value of all the words to the left of the word for a sport that involves a \u2018wicket\u2019 and multiply this value by the total value of all the words above the word for a sport that takes place in a \u2018ring\u2019.\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench", |
| "sample_id": "19", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench - Sample 73", |
| "description": "Task from inspect_evals/zerobench dataset, sample 73", |
| "prompt": "The girls in the picture have different degrees of makeup and hairstyling. The time it takes each to complete their makeup is as follows:\n1.\tShort hair +5 min / Long hair down +10 min / Long hair styled +20 min\n2.\tEyeliner +10 min\n3.\tBlush +3 min\n4.\tLipstick +2 min\n5.\tAccessories +5 min\n6.\tEyeshadow Total time * 2\n\nFor each girl in the picture, first calculate how long in minutes it takes them to finish their makeup. Then, compute the sum of all the times.\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench", |
| "sample_id": "73", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench - Sample 21", |
| "description": "Task from inspect_evals/zerobench dataset, sample 21", |
| "prompt": "These images show books in a library. A book is considered out of place if, starting from the left, it does not follow naturally from the previous book. For each image, identify the out-of-place books. If a book is identified as out of place, it is removed, and the next book is then compared relative to the previous book, disregarding the removed book.\n\nYour final answer is the sum of the years of all the in-order books minus the sum of the years of all the out-of-place books across all images.\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench", |
| "sample_id": "21", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench - Sample 70", |
| "description": "Task from inspect_evals/zerobench dataset, sample 70", |
| "prompt": "The length of the pen is 12 cm.\n(1) Find the length of the snake from the tip of the tail to the tip of the tongue (to the nearest cm)\n(2) Find the width of the snake at the thickest part of the snake (to the nearest cm)\n\nAssuming the picture is flat (no distortion, or from the angle of the camera). Assume the snake is a cylindrical body with the length and width found earlier. What is the volume of the snake in cm^3 to 2dp?\n\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench", |
| "sample_id": "70", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench_subquestions - Sample 82_1", |
| "description": "Task from inspect_evals/zerobench_subquestions dataset, sample 82_1", |
| "prompt": "How many books are there?\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench_subquestions", |
| "sample_id": "82_1", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench_subquestions - Sample 53_1", |
| "description": "Task from inspect_evals/zerobench_subquestions dataset, sample 53_1", |
| "prompt": "You are trying to crack a safe, and you need a 6-digit code. You find the following clues and items left behind by the owner of the safe. What are the first 4 digits of the code?\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench_subquestions", |
| "sample_id": "53_1", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench_subquestions - Sample 4_1", |
| "description": "Task from inspect_evals/zerobench_subquestions dataset, sample 4_1", |
| "prompt": "Including both real and cartoon, how many cats are there in the image?\n\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench_subquestions", |
| "sample_id": "4_1", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench_subquestions - Sample 76_2", |
| "description": "Task from inspect_evals/zerobench_subquestions dataset, sample 76_2", |
| "prompt": "There is an item with the description \u201cINMOTION fast charge power bank 10kmAh 20 W\u201d. What is the price of this item?\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench_subquestions", |
| "sample_id": "76_2", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench_subquestions - Sample 84_3", |
| "description": "Task from inspect_evals/zerobench_subquestions dataset, sample 84_3", |
| "prompt": "What is the total number of socks without matching pairs? Exclude the socks labelled \"DONNAY\"\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench_subquestions", |
| "sample_id": "84_3", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench_subquestions - Sample 89_5", |
| "description": "Task from inspect_evals/zerobench_subquestions dataset, sample 89_5", |
| "prompt": "In the image of the drying rack, there is a grey coloured tracksuit bottoms on the top right. Is this in the picture of the folded clothes?\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench_subquestions", |
| "sample_id": "89_5", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench_subquestions - Sample 44_1", |
| "description": "Task from inspect_evals/zerobench_subquestions dataset, sample 44_1", |
| "prompt": "What is the plot number of the house diagonally opposite to the house that is located two houses to the right of Dasthagir & Sons? Give just the plot number, e.g., '8'.\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench_subquestions", |
| "sample_id": "44_1", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench_subquestions - Sample 85_3", |
| "description": "Task from inspect_evals/zerobench_subquestions dataset, sample 85_3", |
| "prompt": "The set of tiles on the left appear completely unbroken, however the mirror in the middle reveals their true appearance, shown on the right. What is the coordinates of the tile with the largest area of cracking? Give your answer in the form '(Row,Column)'.\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench_subquestions", |
| "sample_id": "85_3", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench_subquestions - Sample 30_3", |
| "description": "Task from inspect_evals/zerobench_subquestions dataset, sample 30_3", |
| "prompt": "How many dumbbells are stored in the racks?\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench_subquestions", |
| "sample_id": "30_3", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| }, |
| { |
| "name": "inspect_evals/zerobench_subquestions - Sample 11_2", |
| "description": "Task from inspect_evals/zerobench_subquestions dataset, sample 11_2", |
| "prompt": "Is there an en passant square? If so, what is the square?\n\n\nLet's think step by step and give the final answer in curly braces, like this: {final answer}", |
| "mcp_config": { |
| "inspect": { |
| "command": "docker", |
| "args": [ |
| "run", |
| "--rm", |
| "-i", |
| "--init", |
| "-v", |
| "/var/run/docker.sock:/var/run/docker.sock", |
| "-v", |
| "/home/genteki/.cache/inspect_evals:/root/.cache/inspect_evals", |
| "hud-inspect:latest" |
| ] |
| } |
| }, |
| "setup_tool": { |
| "name": "load_task", |
| "arguments": { |
| "task_name": "inspect_evals/zerobench_subquestions", |
| "sample_id": "11_2", |
| "model": "claude-sonnet-4-20250514" |
| } |
| }, |
| "evaluate_tool": { |
| "name": "score_task", |
| "arguments": {} |
| }, |
| "agent_config": { |
| "agent_class": "server.agents.InspectAgent", |
| "disallowed_tools": [ |
| "_store_agent_message", |
| "load_task", |
| "score_task" |
| ], |
| "allowed_tools": [], |
| "initial_screenshot": false |
| }, |
| "system_prompt": "You are a helpful AI assistant. Follow the instructions in the user messages carefully." |
| } |
| ] |
