---
title: README
emoji: 🏃
colorFrom: pink
colorTo: blue
sdk: static
pinned: false
---

# Human Behavior Atlas and OmniSapiens Resources

This repository hosts the official models and datasets released as part of the Human Behavior Atlas (HBA) project — a benchmark and foundation model ecosystem for unified multimodal social behavior understanding and reasoning. Resources include unified OmniSapiens foundation models, specialized behavioral adapters, and Human Behavior Atlas benchmark datasets.

---

---

# Models

| Model | Description |
|---|---|
| [OmniSapiens 2.0](https://huggingface.co/HumanBehaviorAtlas/OmniSapiens2.0) | Latest and strongest OmniSapiens reasoning foundation model trained with HARPO for unified multimodal social behavior understanding. |
| [OmniSapiens SFT](https://huggingface.co/HumanBehaviorAtlas/omnisapiens_sft) | Supervised fine-tuned OmniSapiens model used as the initialization model for RL training. |
| [BAM — Humour Detection](https://huggingface.co/HumanBehaviorAtlas/omnisapiens_bam_humour_detection) | Behavioral Adapter Module specialized for multimodal humor detection. |
| [BAM — Sarcasm Detection](https://huggingface.co/HumanBehaviorAtlas/omnisapiens_bam_sarcasm_detection) | Behavioral Adapter Module specialized for sarcasm understanding. |
| [BAM — Sentiment Polarity (MOSEI)](https://huggingface.co/HumanBehaviorAtlas/omnisapiens_bam_sentiment_polarity_mosei) | Behavioral Adapter Module for multimodal sentiment polarity prediction. |

**OmniSapiens 2.0** is the latest and strongest OmniSapiens foundation model for multimodal social behavior reasoning. The model is trained using **Heterogeneity-Aware Relative Policy Optimization (HARPO)** on the Human Behavior Atlas benchmark, enabling more balanced and consistent learning across diverse behavioral tasks and domains. The model demonstrates strong multi-task performance, out-of-distribution generalization, and more interpretable reasoning traces.

---

# Benchmark

| Dataset | Description |
|---|---|
| [Human Behavior Atlas (Parquet)](https://huggingface.co/datasets/HumanBehaviorAtlas/human_behavior_atlas) | Primary HBA release optimized for Hugging Face `datasets` workflows. |
| [Human Behavior Atlas (TAR)](https://huggingface.co/datasets/HumanBehaviorAtlas/human_behavior_atlas_tar) | TAR-based release for custom preprocessing and large-scale storage workflows. |

The version distributed in **Parquet format**, is generally easier and faster to use with the Hugging Face `datasets` library and most downstream training pipelines. Although, we also provide the tar version to allow for flexible use. 


---

# Citation

If you use Human Behavior Atlas or OmniSapiens in your research, please cite the corresponding work below.

## Human Behavior Atlas

```bibtex
@inproceedings{
ong2026human,
title={Human Behavior Atlas: Benchmarking Unified Psychological And Social Behavior Understanding},
author={Keane Ong and Wei Dai and Carol Li and Dewei Feng and Hengzhi Li and Jingyao Wu and Jiaee Cheong and Rui Mao and Gianmarco Mengaldo and Erik Cambria and Paul Pu Liang},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=ZKE23BBvlQ}
}
```

## OmniSapiens 2.0

```bibtex
@article{ong2026omnisapiens,
  title={Omnisapiens: A foundation model for social behavior processing via heterogeneity-aware relative policy optimization},
  author={Ong, Keane and Boughorbel, Sabri and Xiao, Luwei and Ekbote, Chanakya and Dai, Wei and Qu, Ao and Wu, Jingyao and Mao, Rui and Hoque, Ehsan and Cambria, Erik and others},
  journal={arXiv preprint arXiv:2602.10635},
  year={2026}
}
```