File size: 6,269 Bytes
ee1d0e5
 
 
 
 
 
 
 
 
 
04e6772
ee1d0e5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6dcfd5d
ee1d0e5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
21e76c0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6f4f3fd
21e76c0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
---
pretty_name: Human Behavior Atlas (HBA)
language:
- en
license: other
task_categories:
- text-classification
- audio-classification
- image-classification
- video-classification
- text-generation
tags:
- human-behavior
- social-intelligence
- multimodal
- benchmarking
- psychology
- affective-computing
- emotion-recognition
- intent-recognition
- sarcasm-detection
- depression-detection
- anxiety-detection
- ptsd
- nonverbal-behavior
- pose-estimation
- opensmile
size_categories:
- 100K<n<1M
---

# Human Behavior Atlas (HBA)

Human Behavior Atlas (HBA) is a **unified benchmark for multimodal behavioral understanding**.  
It aggregates and standardizes multiple behavioral datasets into a single training and evaluation framework, enabling consistent training and evaluation of foundation models on psychological and social behavior tasks (e.g., emotion, intent, sarcasm, mental health signals, nonverbal behavior).

- **Dataset on Hugging Face:** https://huggingface.co/datasets/keentomato/human_behavior_atlas  
- **Official GitHub:** https://github.com/MIT-MI/human_behavior_atlas/tree/main  
- **Paper (ICLR 2026 Main Conference):** https://openreview.net/forum?id=ZKE23BBvlQ

---

## What’s inside

When you download the dataset, you will find:

### 1) JSONL splits (centralized indices)
These JSONL files sit at the **root level** and define all benchmark samples:
- `final_v8_train_cleaned_2.jsonl` — training set  
- `final_v8_val_cleaned.jsonl` — validation set  
- `final_v8_test_cleaned.jsonl` — test set  

Each line is a **self-contained sample** that provides:
- the **prompt** (problem statement),
- the **ground-truth** answer/label,
- pointers to any **raw media** (video/audio/text/image),
- pointers to any **pre-extracted behavioral features** (`.pt`).
- miscellaneous information such as the dataset, task, class label, modalities present (i.e. modality signature)

### 2) Raw media files
Subdirectories contain the raw media referenced by each JSONL sample:
- video / audio files
- (optional) associated text files

### 3) Behavioral feature files (`.pt`)
Pre-extracted features for common behavioral signals, such as:
- pose features (video)
- OpenSMILE features (audio)
- (and other feature types included in the release)

---

## Data format

HBA uses JSONL to provide a **unified sample schema** across heterogeneous source datasets.  
All file paths are **relative to the dataset root** (the same directory as the JSONL files).

### Example sample
```json
{
  "problem": "<audio>\nDon't forget a jacket.\nThe above is a speech recording along with the transcript from a clinical context. What emotion is the speaker expressing? Answer with one word from the following: anger, disgust, fear, happy, neutral, sad",
  "answer": "sad",
  "images": [],
  "videos": [],
  "audios": ["cremad_dataset_audio/1077_DFA_SAD_XX.wav"],
  "dataset": "cremad",
  "texts": [],
  "modality_signature": "text_audio",
  "ext_video_feats": ["pose/cremad_dataset_audio/1077_DFA_SAD_XX.pt"],
  "ext_audio_feats": ["opensmile/cremad_dataset_audio/1077_DFA_SAD_XX.pt"],
  "task": "emotion_cls",
  "class_label": "sad"
}
```

---

# Field-Level Explanation

### Key fields:

- **problem / answer** — the prompt and ground-truth label  
- **images, videos, audios, texts** — relative paths to raw media files  
- **ext_video_feats, ext_audio_feats** — relative paths to pre-extracted behavioral feature files (.pt)  
- **modality_signature** — indicates which modalities are present (e.g., text_audio, video, text_video)  
- **dataset** — source dataset name  
- **task / class_label** — behavioral task type and label  

### Detailed explanation:

- `problem`: The full prompt presented to the model. May contain modality markers such as `<audio>` or `<video>` and includes the task instruction.
- `answer`: The ground-truth label expected during evaluation.
- `images`, `videos`, `audios`, `texts`: Lists of relative paths to raw media files stored in subdirectories under the dataset root. Empty lists indicate that the modality is not present.
- `ext_video_feats`, `ext_audio_feats`: Lists of relative paths to pre-extracted behavioral feature tensors (`.pt` files), also stored in subdirectories under the dataset root.
- `modality_signature`: A compact indicator of which modalities are present for the sample (e.g., `text_audio`, `video`, `text_video`).
- `dataset`: The original source dataset name (e.g., `cremad`), enabling provenance tracking.
- `task`: High-level behavioral task identifier (e.g., `emotion_cls`).
- `class_label`: Canonical ground-truth class associated with the sample.

The dataloader uses this JSONL as the centralized index to locate and load all raw media and feature files for each sample into the model.

---

# Data Loading Workflow

1. Read a JSONL entry.
2. Parse `problem`, `task`, and `class_label`.
3. Load raw media using relative paths.
4. Optionally load `.pt` behavioral feature tensors.
5. Construct the multimodal sample for training or evaluation.

This unified indexing structure enables heterogeneous behavioral datasets to be standardized under a single multimodal evaluation framework.

---

# Instructions to Download the Full Benchmark (JSONLs + Raw Media)

The dataset consists of:

- JSONL split files at the repository root
- Multipart tar archive under `parts/`

Files:

```
parts/human_behaviour_data.tar.part-000
...
parts/human_behaviour_data.tar.part-009
```

You must download **both** the JSONLs and the archive parts.

If you have sufficient disk space, the simplest and safest approach is to use huggingface-cli:

```bash
pip install -U "huggingface_hub[cli]"
huggingface-cli login

huggingface-cli download keentomato/human_behavior_atlas \
  --repo-type dataset \
  --local-dir hba_download \
  --local-dir-use-symlinks False
```

Then merge and extract:

```bash
cd hba_download
cat parts/human_behaviour_data.tar.part-* > human_behaviour_data.tar
tar -xf human_behaviour_data.tar
```

After extraction, your directory structure will look like:

```
hba_download/
├── final_v8_train_cleaned_2.jsonl
├── final_v8_val_cleaned.jsonl
├── final_v8_test_cleaned.jsonl
├── pose/
├── opensmile/
├── cremad_dataset_audio/
├── ...
```