Dataset Viewer
Auto-converted to Parquet Duplicate
prompt
listlengths
1
1
data_source
stringclasses
1 value
extra_info
dict
[ { "content": "x", "role": "user" } ]
alignx
{ "chosen": "How much is the house worth? How big is it? If you have enough equity you might want to consider selling the house and downgrading to one you can buy outright with the proceeds, if such a house exists in your market.\n\nYou're gonna need that 401k when you can no longer work.\n", "demographic": "This p...
[ { "content": "x", "role": "user" } ]
alignx
{ "chosen": "Which state does she live in, which state is her branch located out of?\n", "demographic": "A detail-oriented individual who values straightforward communication and expects clear answers to their questions. This person prefers to gather information directly related to their query and avoids unnecessar...
[ { "content": "x", "role": "user" } ]
alignx
{ "chosen": "Both of my Sisters In Law are still deeply ashamed of shitting all over the bed, nurses and baby when they were given these things. Their kids were born 3+ years ago. \nApparently this is normal, as Mother In Law and my own Mum have id the same thing. My mum finds it funny rather than embarrassing.\n",...
[ { "content": "x", "role": "user" } ]
alignx
{ "chosen": "Alice hasn't been a boy's name since the 18th century and it really wasn't then either. Vincent \"Alice\" Cooper picked the name as part of his performance art.\n\nThat's definitely a male fail, sweetie. Congrats!\n", "demographic": "A detail-oriented professional who adheres closely to traditional g...
[ { "content": "x", "role": "user" } ]
alignx
{ "chosen": "That's exactly how I feel. I'm keeping my legal name of Alex since it's andro, but my friends gave me the nickname Lexi which I really like! But I feel like I'm not allowed to use it when I'm not fem enough 😣\n", "demographic": "This person dislikes parenting.\nA young, trans woman in the early stages...
[ { "content": "x", "role": "user" } ]
alignx
{ "chosen": "I have my step dads last name hyphenated as part of my last name. So basically, my biological father’s last name is part one, and my step dad’s is part two. Having a long name like that can be a bit annoying to be honest (I don’t mind cause it like having both names) but it’s admittedly a hassle sometime...
[ { "content": "x", "role": "user" } ]
alignx
{ "chosen": "* Chocolate sorbet with Havana Club 7 year dark rum\n\n* Tea, all kinds of tea in my favourite teapot\n\n* Indian food \n\n* Sriracha and cream cheese on toast\n\n* Raspberries\n", "demographic": "This person likes gourmet food.\nA young, creative professional who enjoys experimenting with different cu...
[ { "content": "x", "role": "user" } ]
alignx
{ "chosen": "I don't use autopay because I'm sketched about companies pulling money from my bank. Instead I send payments from my bank to the cc company about a week after I get the statement.\n", "demographic": "This person dislikes finance.\nA financially cautious and detail-oriented individual who prefers to man...
[ { "content": "x", "role": "user" } ]
alignx
{ "chosen": "Following the bot's definition I don't believe in Western Society there is such a thing as rape culture. I believe however there are certain sub-cultures or groups where a rape culture may exist e.g. prison, and certain rituals which those conducting may not see as rape, e.g. hazing in the military and s...
[ { "content": "x", "role": "user" } ]
alignx
{ "chosen": "I'm am Intern who's been lucky enough to look after a few if the PRINT (pre-internship rotation) medical students who will be working at my hospital next year. Apart from putting some effort in even if it slows down my work day a bit so they can learn my hospitals idiosyncrasies I've encouraged them to b...
[ { "content": "x", "role": "user" } ]
alignx
{ "chosen": "I like the brushes that have the bristles molded in plastic. If I use a brush that has the bristles twisted in wire I end up scratching the heck out of the bottom of my flask.\n", "demographic": "An individual who is meticulous about their laboratory equipment and prefers straightforward, efficient so...
[ { "content": "x", "role": "user" } ]
alignx
{ "chosen": "In my city folks set up an illegal safe-injection site in a park to deal with the opioid crisis that the city was ignoring. Saved a lot of people from overdoses & disease, and put so much pressure on the city that they had to fast-track their own safe-injection sites. Liberals have been supportive of...
[ { "content": "x", "role": "user" } ]
alignx
{ "chosen": "I would just say everyone’s parents sets the rules for their kids. It’s okay for my kids to say this word, but you do what your parents tell you. That’s usually along the lines of what I tell my students when they get into those sorts of debacles in my classroom. You want to be careful not to undermine t...
[ { "content": "x", "role": "user" } ]
alignx
{ "chosen": "I did also wait for the 2 direct debits to go out once before transferring\n", "demographic": "This person likes finance.\nAn individual who prefers to follow detailed instructions and wait for confirmations before completing financial tasks, ensuring everything is set up correctly before making any si...
[ { "content": "x", "role": "user" } ]
alignx
{ "chosen": "You should've offered to sell it to the mom for double the price.\n", "demographic": "This person likes gaming, culture.\nThis person dislikes parenting.\nA mature gamer who is passionate about the gaming industry and its cultural significance. They advocate for the recognition of video games as a legi...
End of preview. Expand in Data Studio

SOUL

SOUL is the data suite for human behavior simulation used in Reinforcing Human Behavior Simulation via Verbal Feedback (DITTO), spanning conversation, social simulation, social cognition, role-play, and human-centric evaluation.

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:

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

@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}
}
Downloads last month
-

Paper for sunweiwei/Soul