README.md

---
license: apache-2.0
task_categories:
- text-generation
tags:
- human-simulation
- role-play
- social-intelligence
configs:
  - config_name: alignx_train
    default: true
    data_files:
      - split: train
        path: train/alignx_rl_8k.parquet
  - config_name: behaviorchain_train
    data_files:
      - split: train
        path: train/behaviorchain_rl_train.parquet
  - config_name: coser_train
    data_files:
      - split: train
        path: train/coser_rl_train.parquet
  - config_name: fantom_train
    data_files:
      - split: train
        path: train/fantom_rl_train.parquet
  - config_name: hitom_train
    data_files:
      - split: train
        path: train/hitom_rl_train.parquet
  - config_name: humanllm_train
    data_files:
      - split: train
        path: train/humanllm_rl_train.parquet
  - config_name: humanual_book_train
    data_files:
      - split: train
        path: train/humanual_rl_book.parquet
  - config_name: humanual_chat_train
    data_files:
      - split: train
        path: train/humanual_rl_chat.parquet
  - config_name: humanual_email_train
    data_files:
      - split: train
        path: train/humanual_rl_email.parquet
  - config_name: humanual_news_train
    data_files:
      - split: train
        path: train/humanual_rl_news.parquet
  - config_name: humanual_opinion_train
    data_files:
      - split: train
        path: train/humanual_rl_opinion.parquet
  - config_name: humanual_politics_train
    data_files:
      - split: train
        path: train/humanual_rl_politics.parquet
  - config_name: lifechoices_train
    data_files:
      - split: train
        path: train/lifechoices_hard_rl.parquet
  - config_name: mirrorbench_train
    data_files:
      - split: train
        path: train/mirrorbench_rl_train.parquet
  - config_name: mistakes_train
    data_files:
      - split: train
        path: train/mistakes_rl_train.parquet
  - config_name: paratomi_train
    data_files:
      - split: train
        path: train/paratomi_rl_train.parquet
  - config_name: sim_doc_train
    data_files:
      - split: train
        path: train/sim_doc_rl.parquet
  - config_name: sim_math_train
    data_files:
      - split: train
        path: train/sim_math_rl.parquet
  - config_name: social_r1_train
    data_files:
      - split: train
        path: train/social_r1_rl.parquet
  - config_name: socsci210_train
    data_files:
      - split: train
        path: train/socsci210_rl_2k.parquet
  - config_name: sotopia_train
    data_files:
      - split: train
        path: train/sotopia_clean_rl.parquet
  - config_name: twinvoice_train
    data_files:
      - split: train
        path: train/twinvoice_rl_train.parquet
  - config_name: userllm_train
    data_files:
      - split: train
        path: train/userllm_rl_train.parquet
  - config_name: alignx_arbitrary_test
    data_files:
      - split: test
        path: test/alignx_arbitrary_val.parquet
  - config_name: alignx_demo_test
    data_files:
      - split: test
        path: test/alignx_demo_val.parquet
  - config_name: alignx_history16_test
    data_files:
      - split: test
        path: test/alignx_history16_val.parquet
  - config_name: alignx_pair_test
    data_files:
      - split: test
        path: test/alignx_pair_val.parquet
  - config_name: alignx_ugc_test
    data_files:
      - split: test
        path: test/alignx_ugc_val.parquet
  - config_name: behaviorchain_test
    data_files:
      - split: test
        path: test/behaviorchain_val.parquet
  - config_name: coser_test
    data_files:
      - split: test
        path: test/coser_val.parquet
  - config_name: fantom_test
    data_files:
      - split: test
        path: test/fantom_val.parquet
  - config_name: hitom_test
    data_files:
      - split: test
        path: test/hitom_val.parquet
  - config_name: humanllm_test
    data_files:
      - split: test
        path: test/humanllm_val.parquet
  - config_name: humanual_book_test
    data_files:
      - split: test
        path: test/humanual_book_val.parquet
  - config_name: humanual_chat_test
    data_files:
      - split: test
        path: test/humanual_chat_val.parquet
  - config_name: humanual_email_test
    data_files:
      - split: test
        path: test/humanual_email_val.parquet
  - config_name: humanual_news_test
    data_files:
      - split: test
        path: test/humanual_news_val.parquet
  - config_name: humanual_opinion_test
    data_files:
      - split: test
        path: test/humanual_opinion_val.parquet
  - config_name: humanual_politics_test
    data_files:
      - split: test
        path: test/humanual_politics_val.parquet
  - config_name: lifechoices_test
    data_files:
      - split: test
        path: test/lifechoices_val.parquet
  - config_name: mirrorbench_test
    data_files:
      - split: test
        path: test/mirrorbench_val.parquet
  - config_name: mistakes_test
    data_files:
      - split: test
        path: test/mistakes_val.parquet
  - config_name: paratomi_test
    data_files:
      - split: test
        path: test/paratomi_val.parquet
  - config_name: sim_doc_test
    data_files:
      - split: test
        path: test/sim_doc_val.parquet
  - config_name: sim_math_test
    data_files:
      - split: test
        path: test/sim_math_val.parquet
  - config_name: social_r1_test
    data_files:
      - split: test
        path: test/social_r1_val.parquet
  - config_name: socsci210_test
    data_files:
      - split: test
        path: test/socsci210_val.parquet
  - config_name: sotopia_test
    data_files:
      - split: test
        path: test/sotopia_hard_val.parquet
  - config_name: twinvoice_test
    data_files:
      - split: test
        path: test/twinvoice_val.parquet
  - config_name: userllm_test
    data_files:
      - split: test
        path: test/userllm_val.parquet
---

# SOUL

**SOUL** is the data suite for human behavior simulation used in
[Reinforcing Human Behavior Simulation via Verbal Feedback](https://arxiv.org/abs/2605.20506) (DITTO),
spanning conversation, social simulation, social cognition, role-play, and human-centric evaluation.

- 📄 Paper: https://arxiv.org/abs/2605.20506
- 💻 Code: https://github.com/sunnweiwei/OdysSim
- 🤖 Model: https://huggingface.co/sunweiwei/Ditto-8B

## Structure

Files are kept in their original format, bound to our training/eval framework, and are organized
into two folders:

- `train/` — RL training data, one parquet per task.
- `test/` — evaluation data, one parquet per task.

Because each task's `extra_info` schema differs across tasks and between train/eval, each
task+split is exposed as its own config (`<task>_train`, `<task>_test`) so it loads cleanly:

```python
from datasets import load_dataset
train = load_dataset("sunweiwei/Soul", "userllm_train", split="train")
test  = load_dataset("sunweiwei/Soul", "userllm_test",  split="test")
```

You can also read any parquet directly, e.g. `train/userllm_rl_train.parquet` /
`test/userllm_val.parquet`.

## Citation

```bibtex
@article{sun2026ditto,
  title         = {Reinforcing Human Behavior Simulation via Verbal Feedback},
  author        = {Sun, Weiwei and Zhou, Xuhui and Liu, Jiarui and Du, Weihua and Sun, Haojia and Xie, Yiqing and Ma, Qianou and Chen, Sihao and Wan, Mengting and Yang, Longqi and Zhou, Pei and Wu, Sherry and Welleck, Sean and Neubig, Graham and Yang, Yiming and Sap, Maarten},
  year          = {2026},
  eprint        = {2605.20506},
  archivePrefix = {arXiv},
  url           = {http://arxiv.org/abs/2605.20506}
}
```