File size: 3,134 Bytes
7a93477
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
---
license: cc-by-4.0
language:
- en
pretty_name: FACTOR-Bench
size_categories:
- 1K<n<10K
tags:
- benchmark
- vision-language
- clip
- compositionality
- negation
- boolean-operators
- image-text-matching
configs:
- config_name: default
  data_files:
  - split: test
    path: factor_bench.jsonl
---

# FACTOR-Bench

A diagnostic benchmark that measures whether a vision-language scoring
interface executes **Boolean operators** (negation, conjunction, disjunction,
exclusion, NOR) or merely tracks which concepts are mentioned. It contains
1,695 image/caption-pair samples built from COCO val2017, all OWL-ViT-validated.

Introduced in *Similarity Is Not Logic: Factored Inference for Dual-Encoder
Vision-Language Models* (ICML 2026).

- **Project page:** https://sultanmo.github.io/factored-vlm/
- **Evaluation harness and reference results:** https://github.com/SultanMo/factored-vlm

## Usage

```python
from datasets import load_dataset
ds = load_dataset("sulmo/FACTOR-Bench", split="test")
```

Each sample is a two-alternative forced choice: one image and two captions
that share the same concepts but differ in logical structure. Score both
captions and pick the higher; `correct` gives the answer.

```json
{
  "sample_id": "not_explicit_0000",
  "test_type": "operator",
  "operator": "NOT",
  "image_id": 233825,
  "image": "val2017/000000233825.jpg",
  "caption_a": "an orange",
  "caption_b": "there is no orange",
  "correct": "b",
  "parse_a": {"concepts": [{"text": "orange", "is_negated": false}], "operator": "SINGLE"},
  "parse_b": {"concepts": [{"text": "orange", "is_negated": true}],  "operator": "SINGLE"},
  "meta": {"difficulty": "easy", "position_swapped": true}
}
```

Oracle parses are embedded in every sample, so evaluation does not depend on
any text parser. Filter by `test_type` for the three splits:

- **operator** (1,100): NOT / AND / OR / BUT_NOT / NEITHER. The main accuracy metric, chance = 50%.
- **equivalence** (450): De Morgan / double negation / commutativity, `correct="both"`. Metric = score-consistency violation rate.
- **compound** (145): 3-4 concepts, mixed polarity.

## Images

The dataset distributes captions, labels, parses, and **COCO image IDs**, not
the images themselves. Download COCO val2017 (~1 GB) and resolve the `image`
field against it:

```bash
wget http://images.cocodataset.org/zips/val2017.zip && unzip val2017.zip -d coco
```

## Anti-shortcut design

Balanced answer positions (a/b about 50/50), balanced polarity on negation
(about half the correct answers are the negated caption), 5+ template
phrasings per operator, and OWL-ViT validation of every concept presence and
absence.

## License

Annotations: **CC-BY-4.0**. Images: COCO val2017 under its original terms
(not redistributed). See the LICENSE file.

## Citation

```bibtex
@inproceedings{alshehri2026similarity,
  title     = {Similarity Is Not Logic: Factored Inference for Dual-Encoder Vision-Language Models},
  author    = {Alshehri, Sultan and Yang, Zhantao and Zhang, Han and Savvides, Marios},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}
```