Release verified evaluation data and paper training annotations
Browse files- LICENSE +6 -0
- README.md +103 -0
- data/cc3m_qwen.parquet +3 -0
- data/coco_mistral.parquet +3 -0
- data/coco_qwen.parquet +3 -0
- manifest.json +26 -0
LICENSE
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Generated annotations: Creative Commons Attribution 4.0 International (CC BY 4.0).
|
| 2 |
+
Copyright (c) 2026 Bora Kargi, Arnas Uselis, Seong Joon Oh.
|
| 3 |
+
|
| 4 |
+
License terms: https://creativecommons.org/licenses/by/4.0/legalcode
|
| 5 |
+
|
| 6 |
+
This license applies to the authors’ generated annotations. Third-party source captions and images retain their original rights and terms; see README.md.
|
README.md
ADDED
|
@@ -0,0 +1,103 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
pretty_name: Half-Truths
|
| 4 |
+
language:
|
| 5 |
+
- en
|
| 6 |
+
tags:
|
| 7 |
+
- image-text-retrieval
|
| 8 |
+
- compositionality
|
| 9 |
+
- clip
|
| 10 |
+
- benchmark
|
| 11 |
+
- arxiv:2602.23906
|
| 12 |
+
configs:
|
| 13 |
+
- config_name: coco_qwen
|
| 14 |
+
default: true
|
| 15 |
+
data_files:
|
| 16 |
+
- split: test
|
| 17 |
+
path: data/coco_qwen.parquet
|
| 18 |
+
- config_name: coco_mistral
|
| 19 |
+
data_files:
|
| 20 |
+
- split: test
|
| 21 |
+
path: data/coco_mistral.parquet
|
| 22 |
+
- config_name: cc3m_qwen
|
| 23 |
+
data_files:
|
| 24 |
+
- split: test
|
| 25 |
+
path: data/cc3m_qwen.parquet
|
| 26 |
+
---
|
| 27 |
+
|
| 28 |
+
# Half-Truths
|
| 29 |
+
|
| 30 |
+
Evaluation data for **[Half-Truths Break Similarity-Based Retrieval](https://arxiv.org/abs/2602.23906)**, NeurIPS 2026. Bora Kargi, Arnas Uselis, Seong Joon Oh.
|
| 31 |
+
|
| 32 |
+
A half-truth extends a correct image description with an incorrect entity or relation. This release contains the exact 1,437 comparisons used for the paper's verified evaluation.
|
| 33 |
+
|
| 34 |
+
| Configuration | Images from | Foils from | Entity | Relation | Total |
|
| 35 |
+
|---|---|---|---:|---:|---:|
|
| 36 |
+
| `coco_qwen` | COCO | Qwen | 254 | 255 | 509 |
|
| 37 |
+
| `coco_mistral` | COCO | Mistral | 252 | 202 | 454 |
|
| 38 |
+
| `cc3m_qwen` | CC3M | Qwen | 251 | 223 | 474 |
|
| 39 |
+
|
| 40 |
+
These are three evaluation settings. All use the `test` split; they are not training/validation/test partitions of one source. Comparisons may share images, so row counts are not unique image counts. Do not train on these evaluation captions.
|
| 41 |
+
|
| 42 |
+
## Load
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from datasets import load_dataset
|
| 46 |
+
|
| 47 |
+
dataset = load_dataset("kbora/Half-Truths", "coco_qwen", split="test")
|
| 48 |
+
sample = dataset[0]
|
| 49 |
+
image = sample["image"]
|
| 50 |
+
anchor, full_truth, half_truth = (
|
| 51 |
+
sample["anchor"], sample["full_truth"], sample["half_truth"]
|
| 52 |
+
)
|
| 53 |
+
```
|
| 54 |
+
|
| 55 |
+
The Parquet files embed the original image bytes used by evaluation, rather than annotation-tool thumbnails. `image_sha256` identifies each image's contents.
|
| 56 |
+
|
| 57 |
+
## Fields
|
| 58 |
+
|
| 59 |
+
- `id`: stable comparison identifier, including source configuration.
|
| 60 |
+
- `image`: image associated with the comparison.
|
| 61 |
+
- `source_dataset`, `source_image_id`, `foil_generator`: provenance.
|
| 62 |
+
- `image_sha256`: checksum of the embedded image bytes.
|
| 63 |
+
- `original_caption`: source caption.
|
| 64 |
+
- `anchor`: correct shorter description.
|
| 65 |
+
- `full_truth`: correct extended description.
|
| 66 |
+
- `half_truth`: description with an incorrect addition.
|
| 67 |
+
- `triplet_type`: `entity` or `relation`.
|
| 68 |
+
- `foil_type`: the type of edit used to construct the incorrect addition.
|
| 69 |
+
|
| 70 |
+
## Evaluation
|
| 71 |
+
|
| 72 |
+
Normalize image and text embeddings and compute cosine similarity `s`.
|
| 73 |
+
|
| 74 |
+
- **Half-Truth accuracy:** fraction with `s(image, anchor) > s(image, half_truth)`.
|
| 75 |
+
- **Correct-Ordering Rate:** fraction with `s(image, full_truth) > s(image, anchor) > s(image, half_truth)`.
|
| 76 |
+
|
| 77 |
+
Use strict comparisons; ties do not count as correct. Report overall, entity, and relation results for each configuration. Overall accuracy weights comparisons equally. The paper's three-source average weights configurations equally. Both metrics use the same released rows.
|
| 78 |
+
|
| 79 |
+
On `coco_qwen`, the paper reports CLIP 34.0% and CS-CLIP 76.4% Half-Truth accuracy. See the [code repository](https://github.com/kargibora/CS-CLIP) for evaluation and the [training dataset](https://huggingface.co/datasets/kbora/CS-CLIP-Training) for prepared training annotations.
|
| 80 |
+
|
| 81 |
+
## Construction and limitations
|
| 82 |
+
|
| 83 |
+
Text-only generators extract caption units and create plausible foils. Image-grounded automated checks precede manual verification; the retained incorrect additions were checked against their images. The suite covers two image sources and two foil generators, with a limited number of comparisons. Human verification reduces accidental true foils but does not establish that every possible interpretation or visual ambiguity has been eliminated. These results measure the specified additive-caption diagnostic, not unrestricted visual reasoning.
|
| 84 |
+
|
| 85 |
+
`manifest.json` records source annotation hashes and comparison counts. Public fields contain the final evaluated captions; internal annotation notes and machine-specific paths are omitted.
|
| 86 |
+
|
| 87 |
+
## Source assets and licensing
|
| 88 |
+
|
| 89 |
+
The authors’ generated annotations are released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). This grant covers their contributions, not third-party source captions or images.
|
| 90 |
+
|
| 91 |
+
Images originate from [COCO](https://cocodataset.org/) and [Conceptual Captions](https://github.com/google-research-datasets/conceptual-captions). Source images retain their original rights and terms. An annotation license does not relicense source images.
|
| 92 |
+
|
| 93 |
+
## Citation
|
| 94 |
+
|
| 95 |
+
```bibtex
|
| 96 |
+
@inproceedings{kargi2026halftruths,
|
| 97 |
+
title={Half-Truths Break Similarity-Based Retrieval},
|
| 98 |
+
author={Kargi, Bora and Uselis, Arnas and Oh, Seong Joon},
|
| 99 |
+
booktitle={Advances in Neural Information Processing Systems},
|
| 100 |
+
year={2026},
|
| 101 |
+
url={https://arxiv.org/abs/2602.23906}
|
| 102 |
+
}
|
| 103 |
+
```
|
data/cc3m_qwen.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:35f395a33ad4c2ef77942fab55923e3888785c4d9dbad8c7a6ff716677443cf3
|
| 3 |
+
size 43206620
|
data/coco_mistral.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:105a88e0097b58a23064cc949a62ed2bed399b29ae705111d3e25a4fe91f6620
|
| 3 |
+
size 74296470
|
data/coco_qwen.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3f114f34790777eb59777e58c617621c602e50f3320138e92b415793f2f843a2
|
| 3 |
+
size 84084785
|
manifest.json
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"coco": {
|
| 3 |
+
"source_sha256": "dfb9671d39096f53ba0b2fcc8c1587036dc23fa42671ee596c4a397e94c502a8",
|
| 4 |
+
"comparisons": 509,
|
| 5 |
+
"types": {
|
| 6 |
+
"entity": 254,
|
| 7 |
+
"relation": 255
|
| 8 |
+
}
|
| 9 |
+
},
|
| 10 |
+
"mistral": {
|
| 11 |
+
"source_sha256": "d27498c0555f2473004d2befeaea9bcf8d02ed9584df7d64be86bf3d7324027e",
|
| 12 |
+
"comparisons": 454,
|
| 13 |
+
"types": {
|
| 14 |
+
"entity": 252,
|
| 15 |
+
"relation": 202
|
| 16 |
+
}
|
| 17 |
+
},
|
| 18 |
+
"cc3m": {
|
| 19 |
+
"source_sha256": "6d2880057ceb523fbd27082f6d69ada7ebc2ec2e614e2dcd855536de1727ae2f",
|
| 20 |
+
"comparisons": 474,
|
| 21 |
+
"types": {
|
| 22 |
+
"entity": 251,
|
| 23 |
+
"relation": 223
|
| 24 |
+
}
|
| 25 |
+
}
|
| 26 |
+
}
|