|
32 | 32 | )
|
33 | 33 |
|
34 | 34 | IMG_PATH = "src/rai_bench/rai_bench/tool_calling_agent/predefined/images/"
|
35 |
| -true_response_inputs: List[BoolImageTaskInput] = [ |
36 |
| - BoolImageTaskInput( |
37 |
| - question="Is the door on the left from the desk?", |
38 |
| - images_paths=[IMG_PATH + "image_1.jpg"], |
39 |
| - ), |
40 |
| - BoolImageTaskInput( |
41 |
| - question="Is the light on in the room?", |
42 |
| - images_paths=[IMG_PATH + "image_2.jpg"], |
43 |
| - ), |
44 |
| - BoolImageTaskInput( |
45 |
| - question="Do you see the plant?", |
46 |
| - images_paths=[IMG_PATH + "image_2.jpg"], |
47 |
| - ), |
48 |
| - BoolImageTaskInput( |
49 |
| - question="Are there any pictures on the wall?", |
50 |
| - images_paths=[IMG_PATH + "image_3.jpg"], |
51 |
| - ), |
52 |
| - BoolImageTaskInput( |
53 |
| - question="Are there 3 pictures on the wall?", |
54 |
| - images_paths=[IMG_PATH + "image_4.jpg"], |
55 |
| - ), |
56 |
| - BoolImageTaskInput( |
57 |
| - question="Is there a plant behind the rack?", |
58 |
| - images_paths=[IMG_PATH + "image_5.jpg"], |
59 |
| - ), |
60 |
| - BoolImageTaskInput( |
61 |
| - question="Is there a pillow on the armchain?", |
62 |
| - images_paths=[IMG_PATH + "image_7.jpg"], |
63 |
| - ), |
64 |
| -] |
65 |
| -false_response_inputs: List[BoolImageTaskInput] = [ |
66 |
| - BoolImageTaskInput( |
67 |
| - question="Is the door open?", |
68 |
| - images_paths=[IMG_PATH + "image_1.jpg"], |
69 |
| - ), |
70 |
| - BoolImageTaskInput( |
71 |
| - question="Is someone in the room?", |
72 |
| - images_paths=[IMG_PATH + "image_1.jpg"], |
73 |
| - ), |
74 |
| - BoolImageTaskInput( |
75 |
| - question="Do you see the plant?", |
76 |
| - images_paths=[IMG_PATH + "image_3.jpg"], |
77 |
| - ), |
78 |
| - BoolImageTaskInput( |
79 |
| - question="Are there 4 pictures on the wall?", |
80 |
| - images_paths=[IMG_PATH + "image_4.jpg"], |
81 |
| - ), |
82 |
| - BoolImageTaskInput( |
83 |
| - question="Is there a rack on the left from the sofa?", |
84 |
| - images_paths=[IMG_PATH + "image_4.jpg"], |
85 |
| - ), |
86 |
| - BoolImageTaskInput( |
87 |
| - question="Is there a plant on the right from the window?", |
88 |
| - images_paths=[IMG_PATH + "image_6.jpg"], |
89 |
| - ), |
90 |
| - BoolImageTaskInput( |
91 |
| - question="Is there a red pillow on the armchair?", |
92 |
| - images_paths=[IMG_PATH + "image_7.jpg"], |
93 |
| - ), |
94 |
| -] |
95 | 35 | ########## SUBTASKS #################################################################
|
96 | 36 | return_true_subtask = CheckArgsToolCallSubTask(
|
97 | 37 | expected_tool_name="return_bool_response", expected_args={"response": True}
|
@@ -141,10 +81,6 @@ def get_spatial_tasks(
|
141 | 81 | question="is there a TV in the room?",
|
142 | 82 | images_paths=[IMG_PATH + "image_4.jpg"],
|
143 | 83 | ),
|
144 |
| - BoolImageTaskInput( |
145 |
| - question="Is there a pillow on the armchain?", |
146 |
| - images_paths=[IMG_PATH + "image_7.jpg"], |
147 |
| - ), |
148 | 84 | ]
|
149 | 85 |
|
150 | 86 | medium_true_inputs = [
|
@@ -181,6 +117,10 @@ def get_spatial_tasks(
|
181 | 117 | question="Is there a rug under the bed?",
|
182 | 118 | images_paths=[IMG_PATH + "image_2.jpg"],
|
183 | 119 | ),
|
| 120 | + BoolImageTaskInput( |
| 121 | + question="Is there a pillow on the armchain?", |
| 122 | + images_paths=[IMG_PATH + "image_7.jpg"], |
| 123 | + ), |
184 | 124 | ]
|
185 | 125 |
|
186 | 126 | easy_false_inputs = [
|
|
0 commit comments