---
license: mit
language:
- en
pretty_name: BVSR data
tags:
  - ai-scientist
  - scientific-discovery
  - creativity
---

# BVSR data

Paper: [Blind Variation, Selective Retention: Improving Creative Idea Generation in AI Scientists](https://arxiv.org/abs/2611.12184)

Code: [https://github.com/zou-group/BVSR](https://github.com/zou-group/BVSR)

For the selective-retention selector, see [`Xkev/SR-judger`](https://huggingface.co/Xkev/SR-judger).

## Contents

- `bv_results/`: verified scores and creativity labels of the blind-variation and prompting-baseline proposals on 12 scientific tasks, and the information content of each task context (Figure 2, Table 1).
- `sr_data/`: training and evaluation sets of selective retention, with the rankings of every compared method (Figure 3).
- `bvsr_simpletes/`: SimpleTES and SimpleTES + BVSR runs on five tasks (Figure 4).

## Usage

`bash reproduce.sh` in the code repository downloads this dataset to `data/` and reproduces Figures 2-4 and Table 1.

## License

MIT.