---
license: other
license_name: research-use-only
license_link: LICENSE
language:
- en
task_categories:
- video-classification
- audio-classification
- text-classification
tags:
- multimodal
- sarcasm-detection
- humor-detection
- mustard
- ur-funny
- affective-computing
size_categories:
- 1K<n<10K
---

# CTM Affective Benchmarks: MUStARD + UR-FUNNY

Preprocessed multimodal media and text splits for the two affective-reasoning
benchmarks used in the `ctm-ai` `exp_affective` experiments:

- **MUStARD** — multimodal sarcasm detection from TV sitcom clips
  (*Friends*, *The Big Bang Theory*, *The Golden Girls*, *Sarcasmaholics*).
- **UR-FUNNY** — multimodal humor detection from TED talk clips.

This repo bundles the **derived media** (raw clips, muted video streams, and
extracted audio) alongside the **JSON text splits**, so an experiment run can
fetch everything from one place.

## Repository layout

```
mustard/
  mmsd_raw_data/utterances_final/   690 mp4  — raw clips (video + audio)
  mustard_muted_videos/             690 mp4  — video-only (audio stream removed)
  mustard_audios/                   356 mp4  — audio-only, test split only
  mustard_dataset/                    9 json — test split + evaluation subsets
  mustard_smoke3.json                        — 3-example smoke test

urfunny/
  urfunny_videos/                   992 mp4  — raw clips (video + audio)
  urfunny_muted_videos/             992 mp4  — video-only (audio stream removed)
  urfunny_audios/                   992 mp4  — audio-only
  data_raw/                           5 json — test split + evaluation subsets
  urfunny_smoke3.json                        — 3-example smoke test
```

Total: 4,728 files, ~4.55 GB.

### Naming conventions

- MUStARD clip ids look like `1_10004` (non-sarcastic pool) and `2_223`
  (sarcastic pool). Audio files add an `_audio` suffix: `2_223_audio.mp4`.
- UR-FUNNY clip ids are plain integers (`1008`), with audio also suffixed
  `_audio` (`1008_audio.mp4`).
- Media files carry the `.mp4` container throughout, **including the
  audio-only tracks** — they are audio streams in an mp4 container, not video.

### Coverage note

`mustard_audios/` contains **356** files, not 690: audio was extracted only for
the clips in `mustard_dataset/mustard_dataset_test.json` (the 356-example test
split). Those 356 ids are a strict subset of the 690 ids in
`mustard_muted_videos/` and `mmsd_raw_data/utterances_final/`. UR-FUNNY has
full 992/992/992 coverage across all three modality folders.

## Text split schemas

`mustard/mustard_dataset/mustard_dataset_test.json` — a dict keyed by clip id
(356 entries):

| field | description |
|---|---|
| `utterance` | the target utterance text |
| `speaker` | speaker of the target utterance |
| `context` | list of preceding utterances |
| `context_speakers` | speakers for each context utterance |
| `show` | source sitcom |
| `sarcasm` | boolean label |

`urfunny/data_raw/urfunny_dataset_test.json` — a dict keyed by clip id
(992 entries):

| field | description |
|---|---|
| `context_sentences` | list of setup sentences |
| `punchline_sentence` | the punchline |
| `label` | humor label |

The remaining JSON files in each folder are evaluation subsets
(`*_subset_5/6/20/100.json`) and retry/missing-id lists used for partial reruns.

## Usage

```python
from huggingface_hub import snapshot_download

# everything (~4.55 GB)
snapshot_download("lwaekfjlk/ctm-affective", repo_type="dataset")

# just the MUStARD text splits
snapshot_download(
    "lwaekfjlk/ctm-affective",
    repo_type="dataset",
    allow_patterns="mustard/mustard_dataset/*",
)

# just UR-FUNNY audio
snapshot_download(
    "lwaekfjlk/ctm-affective",
    repo_type="dataset",
    allow_patterns="urfunny/urfunny_audios/*",
)
```

There is intentionally no `configs:` block in the card metadata: the JSON files
are id-keyed dicts rather than record lists, so the dataset viewer would not
parse them. Load them with `json.load` directly.

## Provenance and licensing

The clips are **derived from third-party copyrighted footage** (TV sitcom
episodes for MUStARD, TED talks for UR-FUNNY) and are redistributed here only
as preprocessed research artifacts. Use is intended for non-commercial academic
research. Rights to the underlying footage remain with their original owners;
consult the upstream datasets for their terms before redistributing.

Please cite the original datasets:

```bibtex
@inproceedings{castro2019towards,
  title     = {Towards Multimodal Sarcasm Detection (An {\_}Obviously{\_} Perfect Paper)},
  author    = {Castro, Santiago and Hazarika, Devamanyu and P{\'e}rez-Rosas, Ver{\'o}nica
               and Zimmermann, Roger and Mihalcea, Rada and Poria, Soujanya},
  booktitle = {Proceedings of the 57th Annual Meeting of the Association for
               Computational Linguistics (ACL)},
  year      = {2019}
}

@inproceedings{hasan2019urfunny,
  title     = {{UR-FUNNY}: A Multimodal Language Dataset for Understanding Humor},
  author    = {Hasan, Md Kamrul and Rahman, Wasifur and Zadeh, Amir and Zhong, Jianyuan
               and Tanveer, Md Iftekhar and Morency, Louis-Philippe and Hoque, Mohammed (Ehsan)},
  booktitle = {Proceedings of the 2019 Conference on Empirical Methods in Natural
               Language Processing (EMNLP)},
  year      = {2019}
}
```