README.md

---
dataset_info:
  features:
  - name: experiment
    dtype: string
  - name: text
    dtype: string
  - name: RTs
    list: float64
  - name: log_RTs
    list: float64
  splits:
  - name: train
    num_bytes: 159575797
    num_examples: 3818
  - name: test
    num_bytes: 39849490
    num_examples: 945
  download_size: 30076965
  dataset_size: 199425287
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
task_categories:
- text-generation
language:
- en
tags:
- cognitive-science
- psychology
- human-behaviour
- psych-201
size_categories:
- 1K<n<10K
pretty_name: Psych-201 Reaction Time Dataset
---

# Psych-201 Reaction Time Dataset

A curated dataset of human reaction time (RT) measurements extracted from [Psych-201](https://github.com/marcelbinz/Psych-201), a large-scale collection of naturalistic language transcriptions of human psychology experiments. The dataset is designed for modeling individual-level response latency distributions across diverse experimental paradigms. Introduced in [*Small Foundation Models of Human Cognition and Behaviour*](https://arxiv.org/abs/2608.05224) (COLM'26).

## Dataset Summary

| | Train | Test | Total |
|---|---|---|---|
| **Samples (participants)** | 3,810 | 953 | 4,763 |
| **Choice–RT pairs** | 1,328,717 | 332,848 | 1,661,565 |
| **Experiments** | 18 | 18 | 18 |

Each **sample** corresponds to a single experimental participant — specifically, the complete sequence of choice–RT pairs recorded from that participant within a given experiment.


## Preprocessing Pipeline

We curated RT data from Psych-201 via a three-stage filtering pipeline. For full details, see `create_rt_dataset.html`.

### 1. Experiment-level filtering

From 39 experiments containing RT measurements, we excluded two unpublished experiments and removed 18 experiments in which more than five RT entries were invalid (zero, negative, or NaN), yielding 19 candidate experiments.

### 2. Sample-level filtering

Among the 19 remaining experiments, we excluded one experiment (*tessler2018notunreasonable*) in which 86.25% of samples exhibited mismatches between the number of recorded responses and corresponding RT measurements. For the remaining 18 experiments, we removed individual samples containing any invalid RT value. This affected 7 samples across 4 experiments: *anllo2024weird* (1/564), *guenther2020ts* (2/145), *rutledge2023happiness* (3/48,925), and *zika2023traitanxiety* (1/89). After filtering, 53,196 valid samples remained.

### 3. Class rebalancing

The *rutledge2023happiness* experiment comprised over 91% of all remaining samples (48,922 of 53,557). To mitigate this severe class imbalance, we applied stratified subsampling, randomly retaining 1% of its samples (489 participants).

Log-transformed RTs were computed for all samples to accommodate the characteristically right-skewed nature of RT distributions.

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("socius/Psych-201-RT")

# Access splits
train = dataset["train"]
test = dataset["test"]

# Example: inspect one participant's data
sample = train[0]
print(f"Experiment: {sample['experiment']}")
print(f"Number of trials: {len(sample['RTs'])}")
print(f"Mean RT: {sum(sample['RTs']) / len(sample['RTs']):.3f}s")
```

## Source

This dataset is derived from:

> Binz et al. (2025). Psych-201. *[GitHub](https://github.com/marcelbinz/Psych-201)*.

## Citation

If you use this dataset, please cite:

```bibtex
@inproceedings{oh2026smallcogfm,
  title     = {Small Foundation Models of Human Cognition and Behaviour},
  author    = {Oh, Nick and Gobet, Fernand},
  booktitle = {Third Conference on Language Modeling (COLM)},
  year      = {2026},
  note      = {arXiv:2608.05224}
}
```