README.md

---
language:
- en
task_categories:
- question-answering
pretty_name: GSBench
tags:
- benchmark
- bioinformatics
- genomic selection
- agents
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-00000-of-00001.parquet
---

# GSBench

GSBench is a benchmark for evaluating whether AI agents can complete executable genomic selection and genomics-related data analysis tasks from natural-language instructions.

## Benchmark content

GSBench contains 40 tasks covering practical genomics and genomic selection scenarios, including genotype quality control, population structure analysis, genetic parameter estimation, genomic prediction, gene mining, functional characterization, and environmental or phenotypic data analysis.

## Data source

The data used in GSBench are based on datasets reported in Liu et al. (2020), Liu et al. (2021), and Yang et al. (2019). During benchmark construction, we selected task-relevant data subsets, standardized file formats, and performed the necessary preprocessing according to the requirements of different evaluation tasks.

References:

- Liu, H.J., Wang, X., Xiao, Y., Luo, J., Qiao, F., Yang, W., Zhang, R., Meng, Y., Sun, J., Yan, S., Peng, Y., Niu, L., Jian, L., Song, W., Yan, J., Li, C., Zhao, Y., Liu, Y., Warburton, M.L., Zhao, J. and Yan, J. (2020). CUBIC: an atlas of genetic architecture promises directed maize improvement. *Genome Biology*, 21, 20.
- Liu, N., Du, Y., Warburton, M.L., Xiao, Y., Yan, J. and Michael, P. (2021). Phenotypic plasticity contributes to maize adaptation and heterosis. *Molecular Biology and Evolution*, 38, 1262-1275.
- Yang, N., Liu, J., Gao, Q., Gui, S., Chen, L., Yang, L., Huang, J., Deng, T., Luo, J., He, L., Wang, Y., Xu, P., Peng, Y., Shi, Z., Lan, L., Ma, Z., Yang, X., Zhang, Q., Bai, M., Li, S., Li, W., Liu, L., Jackson, D. and Yan, J. (2019). Genome assembly of a tropical maize inbred line provides insights into structural variation and crop improvement. *Nature Genetics*, 51, 1052-1059.

## Repository structure

```text
GSBench/
├── README.md
├── data/
│   └── GSBench.jsonl
├── files/
│   ├── Q001/
│   ├── Q002/
│   ├── ...
│   └── Q040/
└── reference_answer/
    ├── Q001/
    ├── Q002/
    ├── ...
    └── Q040/
```

## Task file

The canonical task file is:

```text
data/GSBench.jsonl
```

Each line is one benchmark task. Example:

```json
{
  "id": "Q001",
  "index": 1,
  "task": "Given a VCF-formatted genotype file located at {files/Q001/raw.vcf.gz}, perform original data quality control (QC). Filter out variants with a missing rate >10% and a MAF <0.05. Convert the QC-filtered data into PLINK binary format (geno_qc.bed/bim/fam). Calculate the first five principal components (PCs) and save the proportion of variance explained by each PC to pca_results.txt.",
  "reference_step": 4,
  "categories": [
    "Data quality control",
    "Population genetic structure analysis"
  ],
  "data_files": [
    "files/Q001/raw.vcf.gz"
  ],
  "reference_answer": "reference_answer/Q001"
}
```

## Fields

| Field | Description |
|---|---|
| `id` | Task identifier in `Qxxx` format, such as `Q001`. |
| `index` | Numeric task index. |
| `task` | Natural-language instruction describing the analysis objective. File paths are written inside `{}`. |
| `reference_step` | Expected number of major analysis steps for the task. |
| `categories` | Task category labels. |
| `data_files` | Input files required by the task. Paths are relative to the repository root. |
| `reference_answer` | Folder containing the reference answer files for the task. |

## Input files

Input files are stored under:

```text
files/Qxxx/
```

For example, input files for `Q001` are stored in:

```text
files/Q001/
```

The paths listed in `data_files` are the files that an AI agent should use to execute the corresponding task.

## Reference answers

Reference answers are stored under:

```text
reference_answer/Qxxx/
```

For example, the reference answer for `Q001` is stored in:

```text
reference_answer/Q001/
```

The reference answer folder may contain output tables, figures, summary files, model metrics, intermediate results, or other files needed to judge whether the task was completed correctly.

## Dataset statistics

- Number of tasks: 40

Category counts:

| Category | Count |
|---|---:|
| Gene mining & Functional characterization | 15 |
| Data quality control | 12 |
| Population genetic structure analysis | 11 |
| Genetic parameter estimation & Genomic prediction | 11 |
| Environmental & Phenotypic data analysis | 6 |

A task may belong to multi categories.