README.md

---
license: apache-2.0
language:
- en
task_categories:
- multiple-choice
tags:
- omni-modal
- multimodal
- robustness
- structural-corruption
- video
- audio
size_categories:
- 10K<n<100K
configs:
- config_name: baseline
  data_files: data/dataset_baseline.jsonl
- config_name: single
  data_files: data/dataset_single.jsonl
- config_name: combined
  data_files: data/dataset_combined.jsonl
---

# SCEval — Modality Fault Lines

Data for **Modality Fault Lines: Structural Corruptions Reveal Fragile Omni-Modal Reasoning** (Findings of EMNLP 2026).

SCEval is a human-verified benchmark for omni-modal robustness. Text, vision, and audio all remain
**present**, but controlled corruptions make the evidence inside a channel unreliable. Each corrupted
item is paired with its clean counterpart at the example level, so clean-to-corrupted comparisons are
made on the same underlying question, answer options, and reference label.

Code, full documentation, and the evaluation protocol:
<https://github.com/Zhaolu-K/ModalityFaultLines-SCEval>

## What is in this repository

```
data/
├─ dataset_baseline.jsonl          273 clean tri-modal questions (the baseline reference)
├─ dataset_single.jsonl         29,144 single-modality corruptions (text / vision / audio)
├─ dataset_combined.jsonl       17,875 bimodal + trimodal joint corruptions
└─ dataset_*_preview.json          first 30 rows of each file, indented and field-trimmed
media/
├─ Social_IQ_single_question/    video + audio, 100 base examples
├─ omnibench/                    image + audio, 77 base examples
└─ valor/                        video + audio, 96 base examples
```

`data/*.jsonl` is one question per line and carries everything needed to evaluate: the question text,
the answer options, the reference answer, the corruption metadata, and direct URLs to the media in
`media/`. The media tree mirrors
`<source>/<sample_id>/<modality>/<operator>/<severity>/<variant>/<file>`.

| Split | Rows | Contents |
|---|---|---|
| `baseline` | 273 | Clean examples. Every corruption is compared against these. |
| `single` | 29,144 | One modality corrupted at a time, 14 operators × severity 10/30/50/70. |
| `combined` | 17,875 | 18 bimodal + 11 trimodal preset conditions. |

## Loading

Metadata only (fast — media stays remote):

```python
from datasets import load_dataset

baseline = load_dataset("KZL96/ModalityFaultLines-SCEval", "baseline", split="train")
single   = load_dataset("KZL96/ModalityFaultLines-SCEval", "single",   split="train")

row = single[0]
print(row["question"], row["options"], row["gold_key"])
print(row["visual_url"], row["audio_url"])   # resolve directly over HTTPS
```

Media, all of it (about 112 GB — pick a subfolder if you do not need everything):

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="KZL96/ModalityFaultLines-SCEval",
    repo_type="dataset",
    allow_patterns=["media/omnibench/**"],   # drop this argument for the full tree
    local_dir="./SCEval",
)
```

## Field reference

| Field | Meaning |
|---|---|
| `row_id`, `question_id`, `sample_id` | Identifiers. `sample_id` ties a corrupted row to its clean counterpart. |
| `dataset` | Source benchmark: `Social_IQ_single_question`, `omnibench`, or `valor`. |
| `question`, `original_question` | Question text after / before text corruption (identical unless the text channel was corrupted). |
| `options`, `gold_key`, `gold_answer_raw` | Answer options `A0`–`A3`, the correct key, and its full text. |
| `visual_url`, `audio_url` | Media actually shown to the model for this row. |
| `original_visual_url`, `original_audio_url` | The clean media for the same example. |
| `corrupted_url` | Whichever channel this row corrupts; equals `visual_url` or `audio_url`. |
| `visual_kind` | `video` or `image`. |
| `corruption_modality` | Which channel is corrupted (`文本` text / `视觉` vision / `音频` audio). |
| `corruption_group`, `corruption_type` | Operator family and the specific operator. |
| `severity_base` | `10`, `30`, `50`, or `70`. |
| `severity_variant` | Distinguishes multiple random realizations at one severity. |
| `is_baseline` | `true` only in the `baseline` split. |
| `condition_id`, `condition_label` | `combined` split only: which joint-corruption preset. |
| `human_understandable`, `human_valid`, `human_must_three_modal` | Human-verification outcomes (`是` yes / `否` no). |

Two conventions matter when reporting numbers. Corrupted rows are only interpretable against their
same-`sample_id` baseline row, so the `baseline` split should be evaluated alongside any corruption
split. And for stochastic operators, the multiple `severity_variant` realizations within one
(model, modality, operator, severity) bucket are aggregated **worst-case**: if any variant is answered
incorrectly, the bucket counts as incorrect.

## Licensing and source media

This repository is released under Apache-2.0. The corrupted media is derived from three existing
benchmarks, and the underlying source material remains subject to their original licenses and terms:
[Social-IQ](https://github.com/A2Zadeh/Social-IQ), [OmniBench](https://m-a-p.ai/OmniBench/), and
[VALOR](https://github.com/TXH-mercury/VALOR). Please cite those datasets in addition to this work.

## Citation

```bibtex
@inproceedings{sceval2026,
  title     = {Modality Fault Lines: Structural Corruptions Reveal Fragile Omni-Modal Reasoning},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}
```