BGI-HangzhouAI/Benchmark_Dataset-Primate_mammal_species_classification

Summary

This dataset provides a benchmark for evaluating the model's capability to identify species-specific genomic signatures across varying evolutionary distances. Using whole-genome assembly data from seven diverse mammals—including four primates (Human, Pan troglodytes, Pongo abelii, Callithrix) and three non-primates (Mouse, Cattle, Sheep)—we designed a multi-class species classification task. To rigorously assess generalization and prevent data leakage from homologous regions, we assigned distinct sets of chromosomes to the training, validation, and testing splits respectively. We sampled non-overlapping sequences across four fixed lengths—1k, 8k, 32k, and 128k—to ensure uniform genomic coverage. This task challenges the model to learn global phylogenetic features rather than memorizing specific genomic locations.

Usage

from datasets import load_dataset
#Download the whole dataset
dataset = load_dataset("BGI-HangzhouAI/Benchmark_Dataset-Primate_mammal_species_classification")
#Download a specific task
task_name = "Primate_mammal_species_classification_1024"
dataset = load_dataset(
    "BGI-HangzhouAI/Benchmark_Dataset-Primate_mammal_species_classification",
    data_files = {
        "train": f"{task_name}/train.jsonl",
        "test": f"{task_name}/test.jsonl",
        "eval": f"{task_name}/eval.jsonl",
    }
)

Detailed Information

Task	Task_name	Input fields	Train Seqs	Validation Seqs	Test Seqs
Primate mammal species classification 1k	Primate_mammal_species_classification_1024	{seq, label}	28,000	7,000	7,000
Primate mammal species classification 8k	Primate_mammal_species_classification_8192	{seq, label}	28,000	7,000	7,000
Primate mammal species classification 32k	Primate_mammal_species_classification_32768	{seq, label}	28,000	7,000	7,000
Primate mammal species classification 128k	Primate_mammal_species_classification_131072	{seq, label}	28,000	7,000	7,000

Dataset Split	Chromosomes
Train	The rest of Autosomes
Test	Chr 5, 7, 9, 10, 11
Validation	Chr 1, 3, 12, 13

Species	Label	Accession Numbers	Source
Human	0	GWHCBHP00000000	CNCB
Pan troglodytes	1	GCA_028858775.3	NCBI
Pongo abelii	2	GCA_002880775.3	NCBI
Callithrix	3	GCA_049354665.1	NCBI
Mouse	4	GCA_964188535.1	NCBI
Cattle	5	GCF_002263795.3	NCBI
Sheep	6	GCF_016772045.2	NCBI

Data processing

1. Data Generation

The dataset was constructed by partitioning genomes at the chromosome level to prevent data leakage.

• Method: Sequences were sampled using a non-overlapping sliding window at fixed lengths (1k, 8k, 32k, and 128k).
• Quantity: 4,000 sequences for training per species. 1,000 sequences each for validation and testing per species.

2. Final format

Datasets are saved in JSONL format. Each record contains the following fields:

• "seq" — The DNA sequence string (A/C/G/T, uppercase).
• "label" — A 7-class species indicator (0-6), representing the corresponding species.
• "chr" — The source chromosome.
• "len" — The fixed length of the sequence (e.g., 1024, 8192).
• "species_name" — The name of the species associated with the label.