---
license: apache-2.0
task_categories:
- text-generation
language:
- en
pretty_name: Human Study
size_categories:
- n<1K
---

# human_study

Free-form human Reddit responses derived from
[`snap-stanford/user_study_annotations`](https://huggingface.co/datasets/snap-stanford/user_study_annotations).
The original Reddit post text comes from the HumanLM authors'
[`reddit_post_dict_testset.json`](https://github.com/zou-group/humanlm/blob/6faaf072b14b3efb4d434237d6d2af13f7a91d00/user_study/data/reddit_post_dict_testset.json).

Each row contains an original Reddit post in `prompt` and the response written
by a human-study participant in `target`. Model responses, generated personas,
comparison judgments, and worker identifiers are intentionally excluded.

## Deduplication and splits

- Source annotation revision: `b19978763f7a1e4fcf71c2facfa65af0136cc425`
- Random seed: `42`
- Exactly one annotation is randomly selected per unique `post_id`.
- `post_id` is unique across the complete derived dataset.
- The 5 most compact selected rows form the few-shot `train`
  split to keep prompts within small model context windows.
- All remaining unique posts form the scored `test` split.
- Train rows: 5
- Test rows: 59

The prompt contract is:

```text
Post:
<original Reddit post>

Response:<SPACE>
```

The `Response:` cue ends with exactly one ASCII space so lm-eval can append
few-shot targets and begin generation at a natural token boundary. The target
is the participant's original free-form response.