lacebench commited on
Commit
511bb71
·
verified ·
1 Parent(s): d723ee6

Upload dataset (incl. keyword dicts)

Browse files
README.md ADDED
@@ -0,0 +1,202 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-4.0
3
+ task_categories:
4
+ - image-to-text
5
+ - visual-question-answering
6
+ tags:
7
+ - Multimodal benchmark
8
+ - Vision-Language Models
9
+ - Compositionality
10
+ - Localism-aware compositionality
11
+ - Multimodal knowledge editing
12
+ configs:
13
+ - config_name: default
14
+ data_files:
15
+ - split: train
16
+ path: data/lace_train.parquet
17
+ - split: test
18
+ path: data/lace_test.parquet
19
+ - config_name: keyword_dict
20
+ data_files:
21
+ - split: train
22
+ path: data/train_keyword_dict.parquet
23
+ - split: test
24
+ path: data/test_keyword_dict.parquet
25
+ ---
26
+
27
+ # LACE-Bench: Localism-Aware Compositionality Evaluation Benchmark for Vision-Language Models
28
+
29
+ > **LACE-Bench** is a benchmark for evaluating *localism-aware compositionality* in vision-language models (VLMs) — the ability to selectively integrate local region-level semantics with global scene-level understanding. It comprises two complementary tasks: **Local & Global Compositional Captioning** and **Multimodal Compositional Knowledge Editing**.
30
+
31
+ ## Dataset Card
32
+
33
+ | Field | Info |
34
+ |---|---|
35
+ | **Tasks** | Local & Global Compositional Captioning, Multimodal Compositional Knowledge Editing |
36
+ | **Modality** | Vision-Language |
37
+ | **Split** | Train (9,874 images) / Test (2,183 images) |
38
+ | **Total** | 12,057 images |
39
+ | **Image Source** | [Visual Genome](https://homes.cs.washington.edu/~ranjay/visualgenome/api.html) |
40
+ | **License** | CC BY 4.0 |
41
+
42
+ <!--
43
+ ## Tasks
44
+
45
+ ### 1. Multi-grained Local and Global Compositional Captioning
46
+
47
+ This task evaluates a model's *static selectivity*: can it simultaneously understand the global scene while identifying and grounding constituent local objects?
48
+
49
+ - **Local captioning**: given an atomic region (a single object marked with a colored bounding box), generate a region-specific caption.
50
+ - **Global captioning**: given a compound region (a group of atomic regions), generate a single coherent caption that integrates all constituent local parts while introducing holistic scene-level context (moods, relations, atmosphere) not present in any individual local caption.
51
+
52
+ Evaluation uses standard captioning metrics (BLEU, ROUGE-1, METEOR) against human-annotated reference captions.
53
+
54
+ ### 2. Multimodal Compositional Knowledge Editing
55
+
56
+ This task evaluates a model's *dynamic robustness*: can it apply a localized counterfactual edit (e.g., replacing *referee* with *spectator*) consistently across region-marked images, while preserving all unrelated global semantics?
57
+
58
+ A multimodal edit request is defined as a tuple `(I, r, ph → ph*)`, where:
59
+ - `I` is the image, `r` is the target region
60
+ - `ph` is the original phrase to be replaced, `ph*` is the counterfactual substitute
61
+
62
+ The model is evaluated on whether it correctly reflects `ph*` in both in-scope regions (edited) and correctly retains all out-of-scope regions (unedited).
63
+
64
+
65
+ ## Intended Use
66
+
67
+ LACE-Bench is designed for:
68
+
69
+ - Evaluating **localism-aware compositionality** — whether VLMs can selectively deploy local and global compositional operations as the task demands
70
+ - Measuring **global binding stability**: how consistently local semantic units of atomic regions bind into global captions
71
+ - Quantifying **cross-scale interference**: the degree to which local counterfactual edits propagate into unintended global semantic regions
72
+ - Benchmarking **fine-tuning strategies** (e.g., LoRA, blur+bbox visual grounding) for compositional captioning -->
73
+
74
+
75
+ ## Data Fields
76
+
77
+ Each record corresponds to one image and contains the following fields:
78
+
79
+ | Field | Type | Description |
80
+ |---|---|---|
81
+ | `image_id` | string | Visual Genome image identifier |
82
+ | `regions` | list[object] | Annotated bounding box regions (atomic) |
83
+ | `narratives` | string | Description of the full image |
84
+ | `keywords` | list[object] | Key noun concepts grounded in WordNet |
85
+ | `relation_centric_regions` | list[object] | Groups of region IDs with a human-written relational annotation |
86
+
87
+ ### `regions`
88
+
89
+ Each atomic region corresponds to a single object marked with a distinct colored bounding box.
90
+
91
+ | Key | Type | Description |
92
+ |---|---|---|
93
+ | `id` | string | Region identifier (`{image_id}_{region_index}`) |
94
+ | `color` | string | Bounding box color used for visual grounding (aqua / yellow / lime / red / blue / orange / magenta) |
95
+ | `x`, `y` | float | Top-left corner coordinates of the bounding box |
96
+ | `width`, `height` | float | Width and height of the bounding box |
97
+ | `captions` | list[object] | Human-annotated region-level captions (see below) |
98
+ | `object_ids` | list[int] | Linked object IDs from Visual Genome |
99
+ | `relationships` | list[object] | Scene graph relationships associated with this region (see below) |
100
+
101
+ **`regions[].captions`**
102
+
103
+ | Key | Type | Description |
104
+ |---|---|---|
105
+ | `caption` | string | Original human-written caption for the region (e.g. `"the tall clock on the street"`) |
106
+ | `counterfactual_caption` | string | Minimally edited caption where one noun is replaced with a plausible but incorrect alternative (e.g. `"the tall dart board on the street"`) |
107
+
108
+ **`regions[].relationships`**
109
+
110
+ | Key | Type | Description |
111
+ |---|---|---|
112
+ | `relationship_id` | int | Visual Genome relationship identifier |
113
+ | `predicate` | string | Relation predicate between subject and object (e.g. `"on"`) |
114
+ | `synsets` | list[string] | WordNet synsets for the predicate (e.g. `["along.r.01"]`) |
115
+ | `subject_id` | int | Visual Genome object ID of the subject |
116
+ | `object_id` | int | Visual Genome object ID of the object |
117
+
118
+ ---
119
+
120
+ ### `keywords`
121
+
122
+ Each entry represents a key noun concept extracted from region captions and grounded in WordNet.
123
+
124
+ | Key | Type | Description |
125
+ |---|---|---|
126
+ | `synset_id` | string | WordNet synset identifier (e.g. `clock.n.01`) |
127
+ | `synonyms` | list[string] | Lemma names belonging to this synset (e.g. `["clock"]`) |
128
+ | `nearest_ancestor` | string | Closest hypernym synset in the WordNet hierarchy (e.g. `timepiece.n.01`) |
129
+ | `supersense` | string | Broad semantic category from WordNet lexicographer files (e.g. `noun.artifact`, `noun.person`) |
130
+ | `counterfactual` | list[object] | Human-annotated counterfactual substitutions for this concept (see below) |
131
+
132
+ **`keywords[].counterfactual`**
133
+
134
+ | Key | Type | Description |
135
+ |---|---|---|
136
+ | `human_annotation` | string | Plausible but incorrect substitute chosen by a human annotator (e.g. `"dart board"`) |
137
+ | `candidate` | list[string] | Candidate substitutions presented to the annotator for selection |
138
+
139
+ > `counterfactual` is empty (`[]`) for concepts where no counterfactual annotation was collected.
140
+
141
+ ---
142
+
143
+ ### `relation_centric_regions`
144
+
145
+ Each entry groups multiple atomic regions and provides a human-written description of the relational context among them.
146
+
147
+ | Key | Type | Description |
148
+ |---|---|---|
149
+ | `human_annotation` | string | Free-form description of the spatial or semantic relationship among the grouped regions (e.g. `"The central clock tower... stands as a focal point against the backdrop of the building's pillars."`) |
150
+ | `region_ids` | list[string] | IDs of the atomic regions involved in this relational group (e.g. `["2358647_0", "2358647_1"]`) |
151
+
152
+
153
+ ## `keyword_dict` Config
154
+
155
+ In addition to the per-image records (`default` config), the dataset ships a `keyword_dict` config that maps each WordNet synset ID to the list of surface phrases observed for that concept across the corresponding split's region captions. These dictionaries are useful for keyword-based lookup, counterfactual phrase matching, and lexical normalization of mentions.
156
+
157
+ **Files**
158
+
159
+ - `data/train_keyword_dict.parquet`
160
+ - `data/test_keyword_dict.parquet`
161
+
162
+ **Schema**
163
+
164
+ | Field | Type | Description |
165
+ |---|---|---|
166
+ | `synset_id` | string | WordNet synset identifier (e.g. `tree.n.01`), matching `keywords[].synset_id` in the `default` config |
167
+ | `phrases` | list[string] | All distinct surface phrases (synonyms, plural forms, modifier-noun variants, casing variants) observed for the synset in that split |
168
+
169
+ **Example rows**
170
+
171
+ | `synset_id` | `phrases` |
172
+ |---|---|
173
+ | `leaf.n.01` | `["leaves", "foliage", "leaf", "banana leaf", "dried leaves", "green leaves", ...]` |
174
+ | `tree.n.01` | `["tree", "trunk", "evergreen tree", "pine trees", "fir tree", ...]` |
175
+ | `bus.n.01` | `["bus", "city bus", "double decker bus", "motorbus", "coach", ...]` |
176
+
177
+ **Loading**
178
+
179
+ ```python
180
+ from datasets import load_dataset
181
+ ds = load_dataset("lacebench/LACE-Bench", "keyword_dict")
182
+ ```
183
+
184
+ ## Copyright, License, and Attribution
185
+
186
+ **Copyright © 2026 Anonymous Authors.**
187
+
188
+ LACE-Bench is released under the [Creative Commons Attribution 4.0 International License (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/). When using, redistributing, or adapting this dataset, acknowledge **the authors** as the source, cite **LACE-Bench**, provide links to the dataset and license, and indicate any changes.
189
+
190
+ **Suggested attribution**
191
+
192
+ > LACE-Bench, Anonymous Authors, 2026. https://huggingface.co/datasets/lacebench/LACE-Bench. Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
193
+
194
+ ## Citation
195
+
196
+ ```bibtex
197
+ @dataset{anonymous2026lacebench,
198
+ title = {LACE-Bench: Localism-Aware Compositionality Evaluation Benchmark for Vision-Language Models},
199
+ author = {Anonymous},
200
+ year = {2026},
201
+ }
202
+ ```
data/lace_test.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:41042c91e0228b5fa4e9b88bbe216294098ef76f5706a955de22e8f73d925fa6
3
+ size 19227949
data/lace_train.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:42d8839758aeeb9cc69c6b7bbe4c7f03783237525ebc522813aae830bfd49366
3
+ size 78457729
data/test_keyword_dict.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1dea3544ab2bfe479181ce8ff64676eed6aa4910d9775ce0a86d652ee0e5d99c
3
+ size 158525
data/train_keyword_dict.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c0b1ca3e681b04bd2ba790b9a5c799aca352eda2960e40d6bdf7d6846f96a079
3
+ size 287982