---
license: cc-by-nc-sa-4.0
language:
- en
pretty_name: TRACE-Bench
size_categories:
- 1K<n<10K
task_categories:
- video-text-to-text
- question-answering
tags:
- affective-computing
- emotion
- video-understanding
- multimodal
- benchmark
configs:
- config_name: task_1_grounded_affect_recognition
  data_files:
  - split: test
    path: data/task_1_grounded_affect_recognition.jsonl
- config_name: task_2_regulation_decoding
  data_files:
  - split: test
    path: data/task_2_regulation_decoding.jsonl
- config_name: task_3_affective_cause_reasoning
  data_files:
  - split: test
    path: data/task_3_affective_cause_reasoning.jsonl
- config_name: task_4_affective_effect_reasoning
  data_files:
  - split: test
    path: data/task_4_affective_effect_reasoning.jsonl
- config_name: task_5_full_chain_reconstruction
  data_files:
  - split: test
    path: data/task_5_full_chain_reconstruction.jsonl
---

# TRACE-Bench

TRACE-Bench evaluates how models trace an emotion in a social video, from its causes to its consequences. Each item asks a model to watch a clip and fill a structured answer about one affective episode: what the subject feels, how it shows, whether the display is regulated, what caused the feeling, and what it changes in the people around.

Paper: [arXiv:2610.11410](https://arxiv.org/abs/2610.11410).

This repository holds the questions, the reference answers and the task instructions. The video clips and their subtitles are in the gated repository [county/TRACE-Bench-media](https://huggingface.co/datasets/county/TRACE-Bench-media).

## Tasks

| Config | Task |
|---|---|
| `task_1_grounded_affect_recognition` | Task 1. Grounded Affect Recognition |
| `task_2_regulation_decoding` | Task 2. Regulation Decoding |
| `task_3_affective_cause_reasoning` | Task 3. Affective Cause Reasoning |
| `task_4_affective_effect_reasoning` | Task 4. Affective Effect Reasoning |
| `task_5_full_chain_reconstruction` | Task 5. Full Chain Reconstruction |

Every task is a single `test` split.

## Files

```
data/
  task_1_grounded_affect_recognition.jsonl
  task_2_regulation_decoding.jsonl
  task_3_affective_cause_reasoning.jsonl
  task_4_affective_effect_reasoning.jsonl
  task_5_full_chain_reconstruction.jsonl
  videos.csv
instructions/
  task_1_grounded_affect_recognition.md
  ...
  task_5_full_chain_reconstruction.md
```

Each line of a task file is one item.

| Field | Meaning |
|---|---|
| `unique_id` | item id, `<video_id>_T<task>_<chain>_<k>` |
| `video_id` | the clip the item is about, a row of `videos.csv` |
| `task_id`, `task_name` | task number and name |
| `chain_id` | the affective chain inside the clip that the item belongs to |
| `known_fields` | the fields given in the question, as a JSON object |
| `question` | the prompt as a list of lines: task statement, focal context, known fields, answer template |
| `answer` | the reference answer, with the fields of the answer template that are populated |

`videos.csv` lists every clip with the collection it comes from: `video_id`, `source_collection`, `has_subtitle`.

`instructions/` holds one task instruction per task. It defines the fields of that task and their admissible values.

## Loading

```python
from datasets import load_dataset

from huggingface_hub import hf_hub_download

task = "task_1_grounded_affect_recognition"
items = load_dataset("county/TRACE-Bench", task, split="test")
instruction = open(hf_hub_download("county/TRACE-Bench", f"instructions/{task}.md", repo_type="dataset")).read()
prompt = instruction.strip() + "\n\n" + "\n".join(items[0]["question"])
```

The prompt refers to the clip `items[0]["video_id"]`. A model receives the clip, its subtitles if it reads text, and the prompt, which is the task instruction followed by the question, and returns a JSON object that follows the answer template.

## Evaluation

Categorical fields are compared directly with the reference, and free-text fields are rated by a judge model, as described in the paper.

## License

The questions and the reference answers are released under [CC BY-NC-SA 4.0](LICENSE). The video clips are not part of this repository. They come from public collections and remain the property of their copyright holders, and they are available for non-commercial research under the terms of [county/TRACE-Bench-media](https://huggingface.co/datasets/county/TRACE-Bench-media).

## Citation

```bibtex
@article{li2026trace,
  title         = {Cognition-Oriented Emotion Tracing from Causes to Consequences in Real-World Social Scenes},
  author        = {Li, Hao and Zhang, Jinye and Li, Bobo and Lee, Mong-Li and Hsu, Wynne and Wang, Zheng and Fei, Hao and Zhang, Min},
  journal       = {arXiv preprint arXiv:2610.11410},
  year          = {2026},
  eprint        = {2610.11410},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI}
}
```

## Contact

sc.lihao@whu.edu.cn