README.md

---
license: cc-by-4.0
---

# SMILE-Next: Teaching Large Language Models to Detect, Classify, and Reason about Laughter

This repository contains the official benchmark dataset for  
**SMILE-Next: Teaching Large Language Models to Detect, Classify, and Reason about Laughter**.

SMILE-Next is a multimodal instruction-following benchmark for laughter understanding. It includes tasks such as laughter detection, laugh-type classification, and reasoning about why laughter occurs.

<img width="960" alt="teaser_camready" src="https://github.com/kaist-ami/SMILE-Next/raw/main/imgs/qualitative_results.png">

## Dataset Splits

SMILE-Next is divided into train, validation, and test splits.

| Split | Number of Samples |
|---|---:|
| Train | 4,766 |
| Validation | 959 |
| Test | 661 |


## Data Sources

Containing youtube source, SMILE-Next contains laughter-related video sources, so please download:

- **SMILE**  
  https://github.com/kaist-ami/SMILE-Dataset

- **UR-FUNNY**  
  https://github.com/ROC-HCI/UR-FUNNY

- **Talkshow_L2L**  
  https://github.com/evonneng/learning2listen  
  We use only the Conan and Fallon videos from Talkshow_L2L. Please refer to the original repository to obtain the corresponding videos.

We provide textual multimodal representations for LLM training. So if you are just training LLM with multimodal textual representation, it is okay not to download the original video.

To access the original videos, users should download them from the corresponding original sources.

The `video_url_or_path` field indicates how to locate each video:

- SMILE-sourced videos:  
  `SMILE/{original_name}`

- UR-FUNNY-sourced videos:  
  `UR-FUNNY/{original_name}`

- Talkshow_L2L-sourced videos:  
  `talkshow_L2L/{original_path}`

For each sample, we provide metadata such as `video_title`, `video_start`, and `video_end` when available. These fields can be used to locate the original clip segment from the corresponding source video.

Some samples may have `null` values for `video_start` and `video_end` when the original source does not provide segment-level timestamps or when the sample is synthetic.

## Data Format

Each split file contains a `data` list. For simple LLM training, you can directly use the `conversations` field. Other fields provide video metadata and multimodal textual representations.

The multimodal textual representation includes information such as:

- `relationship`
- utterances
- visual captions
- acoustic features
- facial action units

```text
data
└── [
    ├── id
    ├── video_url_or_path
    ├── video_title
    ├── video_start
    ├── video_end
    ├── task
    ├── textrep
    │   ├── relationship        optional
    │   ├── "0"
    │   │   ├── utterance / Utterance
    │   │   ├── caption / Video caption
    │   │   ├── acoustic / Acoustic features
    │   │   ├── facial action unit / Facial Action Units
    │   │   └── Speaker          optional
    │   ├── "1"
    │   │   └── ...
    │   └── ...
    └── conversations
        └── [
            ├── { from, value }
            └── { from, value }
        ]
]
```

### Field Descriptions
- `id`: Unique sample identifier.
- `video_url_or_path`: Source video URL or source-specific path.
- `video_title`: Title or identifier of the source video.
- `video_start`: Start timestamp of the clip segment, if available.
- `video_end`: End timestamp of the clip segment, if available.
- `task`: Task type, such as detection, classification, or reasoning.
- `textrep`: Multimodal textual representation of the clip.
- `conversations`: Instruction-following format for LLM training.

### Training and Evaluation
For evaluation, the training should be done on our training sample.
For detailed explanation, please check our github: https://github.com/kaist-ami/SMILE-Next.

**Training**
```
CUDA_VISIBLE_DEVICES=0,1,2,3 FORCE_TORCHRUN=1 llamafactory-cli train llamafactory_configs/qwen25_selfinst_moelora_sft_ds3.yaml

```

**Inference**
```
CUDA_VISIBLE_DEVICES=0 python3 scripts/inference_llama3.py --adapter_name_or_path "./models/saves/llama3-8b/moelora/sft_selfinst" --save_name ./models/saves/llama3-8b/moelora/sft_selfinst/generated_predictions.jsonl
```

## License

SMILE-Next is released under the license specified in this repository.

Please note that SMILE-Next is built from multiple original data sources, including SMILE, UR-FUNNY, Talkshow_L2L, and YouTube-sourced videos. The original videos are not redistributed in this repository. Users are responsible for obtaining the original videos from the corresponding source datasets or platforms and must follow the license, terms of use, and distribution policies of each original source.

This repository provides metadata, instruction data, and textual multimodal representations for research purposes.

## Citation
If you find our code or paper helps, please consider citing:

```bibtex
@inproceedings{jung-mok-etal-2026-smile,
    title = "{SMILE}-Next: Teaching Large Language Models to Detect, Classify, and Reason about Laughter",
    author = "Jung-Mok, Lee  and Sung-Bin, Kim  and Chang, Joohyun  and Hyun, Lee  and Oh, Tae-Hyun",
    editor = "Liakata, Maria  and Moreira, Viviane P.  and Zhang Jiajun  and Jurgens, David",
    booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2026",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.acl-long.2023/",
    pages = "43675--43693",
    ISBN = "979-8-89176-390-6"
}