JobTrends

Given the growing reliance on informal platforms, such as Tele-
gram for job vacancy announcements, traditional approaches to job
title classification face challenges including noisy text, overlapping
semantics between roles, and severe class imbalance. To enable ro-
bust and generalizable modeling across heterogeneous recruitment
sources, we present JobTrends, a large-scale, multi-source dataset
of job postings collected from LinkedIn, Hahu Jobs, and Telegram.
The dataset is released in four complementary versions: (1) a raw
corpus of 381 088 job postings, (2) a cleaned and standardized ver-
sion containing 353 899 entries, (3) a computer science–specific
subset with 55 686 postings, and (4) its manually annotated and
cleaned counterpart. The computer science subset spans 10 occupa-
tional categoriesBy combining data from formal (LinkedIn, Hahu)
and informal (Telegram) sources, JobTrends represents the broad
variability and stylistic range of real-world recruitment language.
To showcase its utility, we describe the collection and organization
process of JobTrends and report baseline results using machine
learning and transformer-based models. While models show vary-
ing strengths across dominant and minority classes, the bench-
marks establish reference points for future work and highlight the
dataset’s potential for advancing research in automated recruitment
analytics.