mlsa-iai-msu-lab/ru_sci_bench_zho_multiclass

RuSciBenchZhoClassification

The RuSciBenchZhoClassification task focuses on predicting the primary semantic category of a scientific article. The dataset used for this benchmark was collected from sciencechina.cn, where articles are indexed with titles, abstracts, and often multiple thematic labels.

Originally, the metadata comprised 90 unique labels. To organize these into coherent groups, we utilized DeepSeek v3.1 for clustering, followed by a manual validation step. Labels that were semantically distant from the main groups were aggregated into an "Other" category.

For the final dataset, we excluded the "Other" category and filtered the data to include only articles associated with a single high-level class. The articles were then split into training and test sets, with class sizes balanced to ensure fair evaluation.

How to evaluate on this task

First, install MTEB version with this task:

pip install git+https://github.com/mlsa-iai-msu-lab/ru_sci_bench_mteb.git@ruscibench

Then run code evaluate a model on this task:

import mteb
from sentence_transformers import SentenceTransformer
model_name = "sentence-transformers/all-MiniLM-L6-v2"
model = mteb.get_model(model_name)
tasks = mteb.get_tasks(tasks=["RuSciBenchZhoClassification"])
results = mteb.evaluate(model, tasks=tasks)