VietJobs: A Vietnamese Job Advertisement Dataset

VietJobs: A Vietnamese Job Advertisement Dataset
Hieu Pham Dinh, Hung Nguyen Huy, Mo El-Haj
College of Engineering and Computer Science, VinUniversity
{24hieu.pd, 25hung.nh, elhaj.m}@vinuni.edu.vn
Abstract
VietJobs is the first large-scale, publicly available corpus of Vietnamese job advertisements, comprising 48,092
postings and over 15 million words collected from all 34 provinces and municipalities across Vietnam. The
dataset provides extensive linguistic and structured information, including job titles, categories, salaries, skills, and
employment conditions, covering 16 occupational domains and multiple employment types (full-time, part-time, and
internship). Designed to support research in natural language processing and labour market analytics, VietJobs
captures substantial linguistic, regional, and socio-economic diversity. We benchmark several generative large
language models (LLMs) on two core tasks: job category classification and salary estimation. Instruction-tuned
models such as Qwen2.5-7B-Instruct and Llama-SEA-LION-v3-8B-IT demonstrate notable gains under few-shot and
fine-tuned settings, while highlighting challenges in multilingual and Vietnamese-specific modelling for structured
labour market prediction. VietJobs establishes a new benchmark for Vietnamese NLP and offers a valuable
foundation for future research on recruitment language, socio-economic representation, and AI-driven labour market
analysis. All code and resources are available at: https://github.com/VinNLP/VietJobs.
Keywords: Vietnamese NLP, job advertisements, large language models, job classification, salary estima-
tion, low-resource language, dataset creation, labour market analysis
1. Introduction
Vietnam’s labour market has expanded alongside
digital transformation, with online recruitment plat-
forms such as TopCV playing an increasingly cen-
tral role in connecting job seekers and employ-
ers. Recent research highlights how such plat-
forms both reflect and reproduce existing

market analysis
1. Introduction
Vietnam’s labour market has expanded alongside
digital transformation, with online recruitment plat-
forms such as TopCV playing an increasingly cen-
tral role in connecting job seekers and employ-
ers. Recent research highlights how such plat-
forms both reflect and reproduce existing social
norms. For instance, explicit references to gender,
age, and physical appearance remain common in
Vietnamese job postings, influencing wage offers
and perceptions of employability (Perroni et al.,
2023; Packard, 2006). Studies have also shown
that job advertisements may privilege particular de-
mographic or aesthetic traits, such as youth or at-
tractiveness, contributing to differentiated opportu-
nities across groups (Perroni et al., 2023). Despite
these observations, the linguistic and structural fea-
tures of Vietnamese recruitment language remain
underexplored, in part due to the limited availability
of large, well-annotated, and publicly accessible
datasets for computational analysis (Tran et al.,
2022).
While previous research on recruitment language
has focused mainly on English and other high-
resource languages, Vietnamese remains compar-
atively under-resourced for NLP. The language’s
tonal structure, compounding morphology, and fre-
quent code-switching with English present addi-
tional challenges for tokenisation, normalisation,
and semantic interpretation (Bonoli and Hinrichs,
2012). This scarcity of data and tools has con-
strained the development of domain-specific and
socially informed NLP applications in Vietnamese.
Recent advances in NLP for Vietnamese, including
work on job description analysis and fake job de-
tection, have begun to address these challenges,
but large, linguistically representative resources
remain limited (Vu et al., 2025; Tran et al., 2022).
This paper introduces VietJobs, the first large-
scale, open-access dataset of Vietnamese job ad-
vertisements. Comprising 48,092 postings col-
lected nationwide,

de-
tection, have begun to address these challenges,
but large, linguistically representative resources
remain limited (Vu et al., 2025; Tran et al., 2022).
This paper introduces VietJobs, the first large-
scale, open-access dataset of Vietnamese job ad-
vertisements. Comprising 48,092 postings col-
lected nationwide, VietJobs integrates linguistic,
demographic, and occupational information to pro-
vide a structured and diverse view of Vietnam’s
online labour market. The dataset supports a wide
range of NLP and analytical studies, including lin-
guistic characterisation of recruitment language
and benchmarking of state-of-the-art large lan-
guage models (LLMs) for job category classification
and salary estimation under zero-shot, few-shot,
and fine-tuned settings (Tran et al., 2022; Vu et al.,
2025; Otani et al., 2024). Through these contribu-
tions, VietJobs establishes a foundational resource
for Vietnamese NLP and labour market analysis.
It bridges computational and socio-economic per-
spectives, offering a platform for studying how lan-
guage represents professional, educational, and
regional variation in recruitment contexts, and sup-
porting future research on language, employment,
and AI applications in Southeast Asia (Perroni et al.,
2023; Bonoli and Hinrichs, 2012; Otani et al., 2024).
2. Related Works
Job advertisements have become a key resource
for computational research on labour markets and
arXiv:2603.05262v1 [cs.CL] 5 Mar 2026

-- 1 of 10 --

recruitment language (Vogt et al., 2023). Advances
in NLP now enable large-scale analyses of job post-
ings to identify hiring trends, examine linguistic pat-
terns, and estimate salaries or job categories (Daw-
son et al., 2020). However, most existing studies fo-
cus on English and other high-resource languages,
leaving low- and mid-resource contexts such as
Vietnamese largely underrepresented.
Several corpora have supported empirical re-
search on recruitment language. The Adzuna
Global Job Listings dataset

ategories (Daw-
son et al., 2020). However, most existing studies fo-
cus on English and other high-resource languages,
leaving low- and mid-resource contexts such as
Vietnamese largely underrepresented.
Several corpora have supported empirical re-
search on recruitment language. The Adzuna
Global Job Listings dataset (Karunarathna, 2025)
contains over 17,000 English job postings with
metadata on compensation and contract type,
while the Djinni Recruitment Dataset (Drushchak
and Romanyshyn, 2024) includes 150,000 jobs
and 230,000 anonymised candidate profiles in En-
glish and Ukrainian. Research on recruitment
language increasingly explores how wording re-
flects social and cultural norms in hiring contexts.
Studies have examined gendered, age-related,
and appearance-based phrasing using approaches
ranging from lexicon-based methods to transformer
models such as RoBERTa (Sharma, 2025), which
perform strongly in identifying recurrent linguistic
patterns in IT and STEM job markets (Kanij et al.,
2024). Work in low-resource languages remains
limited, though multilingual initiatives such as the
AraJobs corpus for Arabic job advertisements (El-
Haj, 2025) highlight the value of culturally grounded
datasets for analysing recruitment discourse.
VietJobs builds upon this direction by introduc-
ing a large-scale, linguistically diverse Vietnamese
dataset that enables analysis of language use and
representation in Southeast Asian labour markets.
Another resource, the Vietnam Jobs Dataset avail-
able on Kaggle (Nguyen, 2025), focuses primarily
on Vietnamese job titles without including full tex-
tual descriptions, which limits its use for NLP-based
analysis. Nevertheless, we employ this dataset for
comparison purposes, as discussed later in Sec-
tion 4.
Salary prediction has also become a key focus
of computational labour market research. Prior
work has applied regression, ensemble learning,
and deep neural models to estimate compensa-
tion, showing that variables such as job

rtheless, we employ this dataset for
comparison purposes, as discussed later in Sec-
tion 4.
Salary prediction has also become a key focus
of computational labour market research. Prior
work has applied regression, ensemble learning,
and deep neural models to estimate compensa-
tion, showing that variables such as job title, com-
pany profile, location, and skills are strong predic-
tors (Pluijmaekers and Lelli, 2022; Bana, 2022).
Many datasets include incomplete or inconsistent
salary information, limiting model reliability (Al-
sheyab et al., 2025; El-Haj, 2025). VietJobs miti-
gates this by including explicit minimum, maximum,
and average salary fields in over 70% of postings,
enabling consistent benchmarking for predictive
and economic analyses.
Overall, VietJobs extends the linguistic and
socio-economic scope of existing Vietnamese
resources and complements regionally focused
datasets such as AraJobs, SkillSpan, JobSkape,
and Djinni (El-Haj, 2025; Zhang et al., 2022; Ma-
gron et al., 2024; Drushchak and Romanyshyn,
2024). By capturing the linguistic and cultural
specificities of Vietnam, it provides a robust re-
source for context-sensitive modelling and con-
tributes to the broader advancement of low-
resource NLP.
3. VietJobs Dataset
3.1. Data Collection and Corpus
Overview
The VietJobs1 dataset was compiled from pub-
licly accessible online recruitment platforms in Viet-
nam during July 2025. Data were collected us-
ing the open-source Crawl4AI framework (Uncle-
Code, 2024), combined with LLM-assisted parsing
through GPT-4o (Hurst et al., 2024) and Gemini 2.5
(Comanici et al., 2025), which enabled structured
extraction from diverse HTML templates while pre-
serving the linguistic integrity of the text. Access to
GPT-4o and Gemini 2.5 was provided through API
integration, enabling scalable and efficient pars-
ing and information extraction. Both models were
guided using a detailed, task-specific prompt to
ensure consistent and schema-compliant outputs.
All

ates while pre-
serving the linguistic integrity of the text. Access to
GPT-4o and Gemini 2.5 was provided through API
integration, enabling scalable and efficient pars-
ing and information extraction. Both models were
guided using a detailed, task-specific prompt to
ensure consistent and schema-compliant outputs.
All collection procedures were conducted ethi-
cally and in compliance with national and institu-
tional regulations, as detailed in Section 7. The
crawling process involved two stages: (1) initial
URL acquisition, which took approximately 4–5
hours, and (2) full-page crawling and LLM-based
information extraction, which required over 7 days.
The resulting corpus captures detailed linguistic
and structural information from job advertisements
representing all 34 provinces and municipalities of
Vietnam, with the highest posting volumes in Hanoi
and Ho Chi Minh City. A statistical overview of the
dataset is presented in Table 1.
With over 15 million words and a vocabulary ex-
ceeding 78,000 unique tokens, VietJobs provides
an extensive and representative sample of contem-
porary Vietnamese recruitment discourse. Its geo-
graphical and occupational diversity supports both
linguistic and computational analyses, enabling the
study of bias, code-switching, and socio-economic
variation across regions and industries. The unified
occupational taxonomy, adapted from ISCO-08,
O*NET, and ESCO standards, ensures analytical
consistency and facilitates downstream applica-
tions such as job classification, salary estimation,
and fairness-aware modelling.
1https://github.com/VinNLP/VietJobs

-- 2 of 10 --

Statistic 	Value
General Information
Total job postings 	48,092
Total tokens (words) 	15,429,581
Average tokens per posting 	321 (mean)
Vocabulary size 	78,002
Proportion of English tokens 	0.32% (49,375 words)
Distinct job categories 	16
Collection duration 	1 week
Geographical Coverage
Regions covered 	34 provinces
Top regions 	Hanoi, Ho Chi Minh City
Salary

job postings 	48,092
Total tokens (words) 	15,429,581
Average tokens per posting 	321 (mean)
Vocabulary size 	78,002
Proportion of English tokens 	0.32% (49,375 words)
Distinct job categories 	16
Collection duration 	1 week
Geographical Coverage
Regions covered 	34 provinces
Top regions 	Hanoi, Ho Chi Minh City
Salary Information
Salary range (min–max) 	1–500M VND/month
Median (min / avg / max) 	10 / 13 / 15M VND
Textual Characteristics
Longest postings 	IT & Digital Engineering (354.8 words)
Shortest postings 	Languages & Translation (275.1 words)
Employment Characteristics
Contract types 	Full-time, Part-time, Internship
Benefits / skills fields 	Limited coverage
Table 1: Summary statistics of the VietJobs
dataset.
3.2. Occupational Category Distribution
The job_category field denotes the occupa-
tional domain or functional area represented in
each advertisement (e.g., Sales, Engineering,
Healthcare). The original categorisation provided
by source platforms consisted of 24 distinct labels,
many of which exhibited semantic overlap or incon-
sistent assignment. For example, positions such
as Financial Analyst and Accountant were placed
under separate categories, despite belonging to
the same broader industrial sector.
To enhance interpretability and comparability,
we conducted a systematic category normalisation
process. All raw category labels were reviewed
and mapped onto a harmonised taxonomy com-
prising 16 consolidated occupational domains. This
approach aligns with established practices in oc-
cupational data standardisation in Vietnam (Mo-
roz and Nguyen, 2019), ensuring analytical con-
sistency while preserving sufficient granularity for
downstream NLP tasks. The resulting taxonomy
maintains a balance between semantic precision
and practical usability, enabling both fine-grained
linguistic analyses and macro-level labour market
comparisons.
Table 2 summarises the distribution of job post-
ings across the 16 categories. The largest seg-
ments correspond to

stream NLP tasks. The resulting taxonomy
maintains a balance between semantic precision
and practical usability, enabling both fine-grained
linguistic analyses and macro-level labour market
comparisons.
Table 2 summarises the distribution of job post-
ings across the 16 categories. The largest seg-
ments correspond to Business, Sales & Customer
Service (8,276 ads) and Manufacturing, Manual
Labour & Mechanics (6,407 ads), reflecting Viet-
nam’s continued emphasis on commerce, produc-
tion, and industrial operations. By contrast, spe-
cialised fields such as Agriculture, Energy & Envi-
ronment (322 ads) and Languages & Translation
(384 ads) account for smaller shares, indicative of
niche professional markets. In total, the dataset
comprises 48,092 valid and categorised job ad-
Job Category (translated to English) 	Ad Count
Business, Sales & Customer Service 	8,276
Manufacturing, Manual Labour & Mechanics 	6,407
Marketing, Communications, Advertising & Content 	5,950
Finance, Accounting, Banking & Insurance 	5,530
Tourism, Hospitality & Services 	4,243
Design, Arts, Entertainment, Media & Journalism 	3,465
Human Resources, Administration, Legal & Consulting 	3,276
Construction, Architecture & Real Estate 	2,811
Information Technology & Digital Engineering 	1,906
Logistics, Transportation & Supply Chain 	1,813
Electrical, Electronics & Telecommunications Engineering 1,236
Education, Training & Research 	1,165
Healthcare, Pharmaceuticals & Biotechnology 	963
Languages & Translation 	384
Agriculture, Energy & Environment 	322
Other Occupations 	345
Total 	48,092
Table 2: Distribution of job advertisements across
the 16 consolidated occupational categories.
vertisements, offering comprehensive coverage of
Vietnam’s online recruitment landscape.
3.3. Salary Distribution
Salary information in the VietJobs dataset
was normalised into three quantitative fields:
salary_min, salary_max, and salary_avg,
each expressed in millions of Vietnamese Dong
(VND) per month. Out of the

.
vertisements, offering comprehensive coverage of
Vietnam’s online recruitment landscape.
3.3. Salary Distribution
Salary information in the VietJobs dataset
was normalised into three quantitative fields:
salary_min, salary_max, and salary_avg,
each expressed in millions of Vietnamese Dong
(VND) per month. Out of the total 48,092 job
postings, 34,365 (71.5%) explicitly specify a salary
range, while the remaining 13,727 (28.5%) are
listed as negotiable. This distribution aligns with
prior research on labour market transparency in
Southeast Asia, where employers often withhold
salary details for competitive or negotiation-related
reasons (Van Thang et al., 2020).
Figure 1 presents the distribution of minimum and
maximum salary values after removing extreme
outliers. The median minimum salary is approxi-
mately 10 million VND, while the median maximum
salary is around 15 million VND. This concentra-
tion suggests that the majority of advertised posi-
tions correspond to entry- or mid-level roles within
Vietnam’s labour market. Nonetheless, the upper
quartiles extend beyond 30 million VND, indicat-
ing the presence of high-paying professional and
managerial roles.
The wider spread of maximum salaries relative to
minimum salaries highlights substantial wage vari-
ability across postings. This variation likely reflects
factors such as job seniority, sectoral demand, and
skill specialisation. For example, service-oriented
occupations—particularly within Business, Sales &
Customer Service—tend to exhibit broader salary
ranges than technical domains like Information
Technology & Digital Engineering. Such disparities
suggest that employers frequently adopt flexible

-- 3 of 10 --

salary ranges to appeal to a wider pool of appli-
cants, a strategy commonly observed in dynamic
and rapidly expanding economies (Lu, 2023).
Figure 1: The distribution of Minimum and Max-
imum Salaries (extreme outliers are removed for
better visualisation)
The salary distribution exhibits a

t flexible

-- 3 of 10 --

salary ranges to appeal to a wider pool of appli-
cants, a strategy commonly observed in dynamic
and rapidly expanding economies (Lu, 2023).
Figure 1: The distribution of Minimum and Max-
imum Salaries (extreme outliers are removed for
better visualisation)
The salary distribution exhibits a right-skewed
pattern, characterised by a concentration of lower-
to mid-income positions alongside a smaller subset
of high-income roles. This structure forms the em-
pirical basis for the salary estimation experiments
described in Section 4, where models are tasked
with predicting salary values for postings using their
key details such as job title or experience required.
4. Evaluating Job Classification and
Salary Estimation with Large
Language Models
4.1. Job Category Classification
The job category classification task focuses on pre-
dicting the standardised job_category label (16
classes) from the textual content of each adver-
tisement, primarily the description field. We
evaluate a range of generative large language mod-
els (LLMs) available on Hugging Face under three
experimental conditions: zero-shot, few-shot, and
fine-tuned. All models employ their chat or in-
struct variants to ensure consistent adherence to
natural language instructions and task formatting.
Prompting and fine-tuning. In the zero-shot
setting, each model receives a concise instruction
prompt requesting a single category prediction from
the 16 predefined occupational labels. This setup
tests a model’s ability to generalise without prior
exposure to task-specific examples. In the few-
shot setting, the prompt includes a small number
of annotated examples before the test instance,
enabling the model to infer task structure and cate-
gory mappings from limited in-context demonstra-
tions. For fine-tuning, the models are trained us-
ing structured instruction–response pairs, where
the prompt contains the job description and the
target output corresponds to the correct

es before the test instance,
enabling the model to infer task structure and cate-
gory mappings from limited in-context demonstra-
tions. For fine-tuning, the models are trained us-
ing structured instruction–response pairs, where
the prompt contains the job description and the
target output corresponds to the correct category
label. This conversational-style format allows mod-
els to internalise consistent mappings between job
descriptions and their respective categories. Dur-
ing inference, model outputs are restricted to the
canonical list of category names. Any responses
that deviate from these standard labels are auto-
matically treated as incorrect. This ensures compa-
rability across models and mitigates the influence
of generative variability, enabling a fair evaluation
of classification accuracy and robustness across
instruction-tuned architectures.
Evaluation metrics. Performance is assessed
using Accuracy and Macro F1 scores on the held-
out test set. Let N denote the total number of sam-
ples, K the number of classes, ˆyi the predicted
label, and yi the ground truth. Accuracy is com-
puted as:
Accuracy = 1
N
N	X
i=1
I(ˆyi = yi)
For each class k, Precision, Recall, and F1 are
defined as:
Precisionk = T Pk
T Pk + F Pk
Recallk = T Pk
T Pk + F Nk
F1k = 2 × Precisionk × Recallk
Precisionk + Recallk
The overall Macro F1 is then given by:
Macro F1 = 1
K
K	X
k=1
F1k
Accuracy provides a measure of overall correctness,
while Macro F1 offers a class-balanced evaluation that
accounts for performance across both frequent and in-
frequent job categories.
4.2. Salary Estimation
The salary estimation task aims to predict the expected
salary range (expressed in the format “X triệu”, which
means X million VND) based on structured job informa-
tion. Each sample includes the following input fields:
job_title, contract_type, location, country,
and experience_required. We evaluate a set of
generative large language models (LLMs) from Hugging-
Face under three configurations:

d in the format “X triệu”, which
means X million VND) based on structured job informa-
tion. Each sample includes the following input fields:
job_title, contract_type, location, country,
and experience_required. We evaluate a set of
generative large language models (LLMs) from Hugging-
Face under three configurations: zero-shot, few-shot and
fine-tuned. All models employ their chat or instruct
variants to enhance instruction-following capability.
Prompting and fine-tuning. In the zero-shot setting,
models are prompted to generate a salary prediction in
the required “X triệu” format using only the provided job

-- 4 of 10 --

attributes. In the few-shot setting, the prompt includes a
small number of example pairs of job attributes and their
corresponding salary values before the test instance,
allowing the model to infer the desired output structure
and relationship through in-context learning. For the fine-
tuned configuration, training instances are structured as
instruction–response pairs following the same conver-
sational schema, enabling models to learn the mapping
between job attributes and corresponding salary values.
During inference, the model output is required to strictly
adhere to the “X triệu” format. The numeric value X
is extracted and compared to the gold-standard salary
value. Predictions that are unparsable or deviate from
the expected format are considered invalid.
Evaluation metrics. Model performance is evaluated
using Root Mean Square Error (RMSE) and the coeffi-
cient of determination (R2) on the held-out test set. Let
N be the number of samples, yi the ground-truth salary,
and ˆyi the predicted value parsed from the model output.
RMSE is computed as:
RMSE =
v
u
u
t 1
N
N	X
i=1
(yi − ˆyi)2
and R2 is defined as:
R2 = 1 −
PN
i=1(yi − ˆyi)2
PN
i=1(yi − ¯y)2
where ¯y denotes the mean of the ground-truth salary
values. RMSE captures the magnitude of prediction er-
rors, while R2 quantifies the proportion of salary variance
explained by the model,

t.
RMSE is computed as:
RMSE =
v
u
u
t 1
N
N	X
i=1
(yi − ˆyi)2
and R2 is defined as:
R2 = 1 −
PN
i=1(yi − ˆyi)2
PN
i=1(yi − ¯y)2
where ¯y denotes the mean of the ground-truth salary
values. RMSE captures the magnitude of prediction er-
rors, while R2 quantifies the proportion of salary variance
explained by the model, providing complementary per-
spectives on predictive accuracy and generalisation.
4.3. Model Selection
To evaluate model performance on Vietnamese job clas-
sification and salary estimation, we selected a diverse
suite of generative large language models (LLMs) that
vary in scale, linguistic coverage, and regional focus.
The selection aims to provide a comprehensive compar-
ison across three categories: globally-trained multilin-
gual models, regionally-oriented ASEAN models, and
Vietnamese-specialised models.
Multilingual models. We include Qwen2.5-7B-
Instruct (Qwen et al., 2025), Llama-3.1-8B-Instruct
(Dubey et al., 2024), Granite-3.3-8B-Instruct (Team,
2025), and Ministral-8B-Instruct-24102. These models
are trained on extensive multilingual corpora and demon-
strate robust generalisation across languages. Their abil-
ity to process code-mixed or English–Vietnamese inputs
makes them suitable for tasks involving international or
bilingual job postings commonly observed in Vietnamese
recruitment data.
ASEAN-focused models. To assess models tailored
for regional linguistic contexts, we evaluate Llama-SEA-
LION-v3-8B-IT (Chan et al., 2024), Sailor2-8B-Chat (Dou
et al., 2025), and SeaLLMs-v3-7B-Chat (Zhang et al.,
2https://huggingface.co/mistralai/
Ministral-8B-Instruct-2410
2024). These models are explicitly designed to improve
coverage of Southeast Asian languages, including Viet-
namese, Thai, and Indonesian. Their training objec-
tives and tokenisation strategies better accommodate
regional linguistic characteristics, which may enhance
performance on language-specific nuances in job adver-
tisements.
Vietnamese-specific models. Finally, we

overage of Southeast Asian languages, including Viet-
namese, Thai, and Indonesian. Their training objec-
tives and tokenisation strategies better accommodate
regional linguistic characteristics, which may enhance
performance on language-specific nuances in job adver-
tisements.
Vietnamese-specific models. Finally, we include
PhoGPT-4B-Chat (Nguyen et al., 2023), BloomVN-
8B-Chat (BlossomsAI, 2025), and Vistral-7B-Chat
(Van Nguyen et al., 2023), all of which are trained or
fine-tuned predominantly on Vietnamese textual data.
These models capture syntactic, lexical, and cultural par-
ticularities of Vietnamese, potentially providing superior
understanding of contextually rich or idiomatic expres-
sions frequently used in job descriptions.
This configuration enables a systematic comparison
between globally multilingual, regionally adapted, and
locally specialised models. By evaluating them under
both zero-shot and fine-tuned settings, we aim to anal-
yse how linguistic coverage, training focus, and cultural
alignment influence LLM performance in Vietnamese
recruitment-related tasks.
4.4. Fine-tuning Configuration
Data Split. The dataset is partitioned into training, de-
velopment, and test subsets using an 80%, 10%, 10%
ratio. The development set is used for hyperparameter
tuning and early stopping, while the final evaluation is
conducted exclusively on the held-out test set to ensure
fair comparison across models.
Fine-tuning Hyperparameters. Fine-tuning is per-
formed on a single NVIDIA A40 GPU using Low-Rank
Adaptation (LoRA) to enable efficient training of large
models. The configuration is consistent across all model
variants to ensure comparability. The key hyperparame-
ters are as follows:
• LoRA parameters: rank = 8, α = 16, dropout = 0.2
• Target modules: q_proj, k_proj, v_proj,
o_proj
• Optimiser: AdamW with learning rate 5 × 10−5
• Batch size: micro-batch size = 4; effective batch
size = 64
• Training epochs: 2
• Maximum sequence length: 512 tokens
•

bility. The key hyperparame-
ters are as follows:
• LoRA parameters: rank = 8, α = 16, dropout = 0.2
• Target modules: q_proj, k_proj, v_proj,
o_proj
• Optimiser: AdamW with learning rate 5 × 10−5
• Batch size: micro-batch size = 4; effective batch
size = 64
• Training epochs: 2
• Maximum sequence length: 512 tokens
• Evaluation and checkpoint frequency: every 200
steps
• Precision: BF16 enabled
Compute and Efficiency. All fine-tuning experiments
were executed under identical computational conditions.
The job classification fine-tuning process required ap-
proximately 5 hours per model, while salary estimation
models converged within 1–2 hours. The use of LoRA
substantially reduced memory consumption and training
time compared to full model fine-tuning, allowing efficient
experimentation across multiple model families.

-- 5 of 10 --

5. Results and Discussion
5.1. Job Classification Results
Table 3 summarises the performance of the evaluated
LLMs on the job classification task across three evalua-
tion settings: zero-shot, few-shot, and fine-tuned.
Zero-shot performance. In the absence of task-
specific examples, Qwen2.5-7B-Instruct achieves the
highest scores, with an accuracy of 0.31 and a Macro F1
of 0.32. This demonstrates its strong cross-lingual gener-
alisation and instruction-following capability, even without
prior exposure to the classification schema. In contrast,
models such as PhoGPT-4B-Chat and Granite-3.3-8B-
Instruct produce inconsistent or malformed outputs that
fail to match the canonical label set, indicating limited
understanding of task constraints and label semantics
under zero-shot prompting.
Few-shot performance. Incorporating a small num-
ber of in-context examples substantially improves per-
formance across all models, with accuracy scores rising
to the 0.4x range. Qwen2.5-7B-Instruct again leads with
0.47 accuracy, followed by Llama-SEA-LION-v3-8B-IT
(0.45) and Sailor2-8B-Chat (0.44). These results high-
light the effectiveness of

rating a small num-
ber of in-context examples substantially improves per-
formance across all models, with accuracy scores rising
to the 0.4x range. Qwen2.5-7B-Instruct again leads with
0.47 accuracy, followed by Llama-SEA-LION-v3-8B-IT
(0.45) and Sailor2-8B-Chat (0.44). These results high-
light the effectiveness of few-shot prompting in structured
text classification, where minimal context allows LLMs
to infer task boundaries, label semantics, and linguistic
cues from examples. The improvement underscores the
adaptability of instruction-tuned models to new domains
with limited supervision.
Fine-tuned performance. Parameter-efficient fine-
tuning yields moderate gains relative to the zero-shot
baseline but does not consistently outperform few-shot
prompting. Accuracy values plateau around 0.3x across
models, suggesting that while fine-tuning enhances task
alignment, it may not fully replicate the contextual rea-
soning and flexibility afforded by in-context learning. No-
tably, Ministral-8B-Instruct-2410 exhibits a slight decline
in both accuracy and Macro F1 after fine-tuning, possibly
reflecting the overfitting problem of the model.
Discussion. Across all experimental settings,
Qwen2.5-7B-Instruct demonstrates the most consistent
and robust performance, achieving the best results in
both zero-shot and few-shot configurations. The strong
results of instruction-tuned multilingual models relative
to Vietnamese-specific ones suggest that large-scale
multilingual pretraining remains advantageous for gen-
eralisable task understanding, even in resource-specific
contexts such as Vietnamese recruitment data.
5.2. Salary Estimation Results
Table 4 summarises the performance of the evaluated
LLMs on the salary estimation task under five experimen-
tal settings: zero-shot, few-shot, fine-tuned on VietJobs,
fine-tuned on the Vietnam Jobs Dataset (Nguyen, 2025),
and fine-tuned on the combined datasets. Overall, model
performance improves consistently as task-specific

mmarises the performance of the evaluated
LLMs on the salary estimation task under five experimen-
tal settings: zero-shot, few-shot, fine-tuned on VietJobs,
fine-tuned on the Vietnam Jobs Dataset (Nguyen, 2025),
and fine-tuned on the combined datasets. Overall, model
performance improves consistently as task-specific su-
pervision is introduced through fine-tuning. For example,
Qwen2.5-7B-Instruct achieves an RMSE of 14.06 and
an R2 of −0.46 in the zero-shot setting on VietJobs,
which improves to an RMSE of 11.32 and an R2 of 0.06
after fine-tuning on both datasets. Comparable trends
are observed across the other models, highlighting the
importance of domain adaptation in salary prediction.
Across all configurations, Llama-SEA-LION-v3-8B-IT
delivers the strongest and most consistent results. Even
without explicit task conditioning, it achieves an RMSE
of 11.72 and an R2 of 0.07 in the zero-shot scenario on
VietJobs, outperforming all other models. With few-shot
prompting, its performance further improves to 10.65
RMSE and 0.16 R2. After fine-tuning on both datasets,
the model maintains robust performance with an RMSE
of 12.40 and R2 of 0.16 when evaluated on the combined
data. This pattern suggests that Llama-SEA-LION effec-
tively integrates pre-trained multilingual knowledge with
task-specific cues, allowing it to generalise well across
both datasets.
The observed performance gains from fine-tuning on
the Vietnam Jobs Dataset and the combined data can
be attributed to increased data diversity and represen-
tativeness. Compared with VietJobs, the Vietnam Jobs
Dataset provides a broader range of industries, salary
brackets, and contextual descriptors, offering richer in-
formation for modelling salary patterns. Fine-tuning on
the combined corpus exposes models to this wider dis-
tribution of job contexts, leading to lower RMSE values
and higher R2 scores—indicative of improved predictive
accuracy and generalisability.
In summary, the results reveal three key

ptors, offering richer in-
formation for modelling salary patterns. Fine-tuning on
the combined corpus exposes models to this wider dis-
tribution of job contexts, leading to lower RMSE values
and higher R2 scores—indicative of improved predictive
accuracy and generalisability.
In summary, the results reveal three key findings: (1)
task-specific fine-tuning substantially enhances model
accuracy for salary estimation; (2) performance generally
follows the trend zero-shot < few-shot < fine-tuned on Vi-
etJobs < fine-tuned on Vietnam Jobs Dataset < fine-tuned
on both datasets; and (3) Llama-SEA-LION-v3-8B-IT
emerges as the most robust and effective model for this
task, demonstrating strong adaptability and generalisa-
tion across different data sources.
6. Limitations
While VietJobs represents the largest publicly available
dataset of Vietnamese job advertisements, several limi-
tations remain. It is sourced from a single online recruit-
ment platform called TopCV, which may not reflect all
sectors or informal employment in Vietnam, resulting in
underrepresentation of some industries or job types. The
postings also follow the platform’s linguistic and structural
conventions, which could introduce systematic biases.
Salary information, though common, is not consis-
tently standardised and may involve rounding or omis-
sion. Textual fields such as job descriptions sometimes
contain duplicated or templated content, potentially af-
fecting linguistic analyses and model outcomes. This
study focuses on two core tasks—job category classifica-
tion and salary estimation—using general-purpose large
language models. These provide initial benchmarks but
do not capture the full scope of possible NLP applica-
tions. Finally, although the dataset excludes identifiable
personal data, downstream users should still apply ap-
propriate ethical and legal safeguards.

-- 6 of 10 --

LLM Job Classification Salary Estimation
Acc Macro F1 RMSE R2
Zero-shot
Qwen2.5-7B-Instruct 0.31 0.32

capture the full scope of possible NLP applica-
tions. Finally, although the dataset excludes identifiable
personal data, downstream users should still apply ap-
propriate ethical and legal safeguards.

-- 6 of 10 --

LLM Job Classification Salary Estimation
Acc Macro F1 RMSE R2
Zero-shot
Qwen2.5-7B-Instruct 0.31 0.32 14.06 -0.46
Llama-3.1-8B-Instruct 0.2 0.2 16.01 -0.87
Ministral-8B-Instruct-2410 0.19 0.16 40.75 -13.12
Llama-SEA-LION-v3-8B-IT 0.26 0.29 11.72 0.07
Sailor2-8B-Chat 0.16 0.16 13.74 -3.29
PhoGPT-4B-Chat 0 0 - -
Granite-3.3-8B-Instruct 0 0 35.37 -8.19
BloomVN-8B-chat 0 0 18.73 -0.91
SeaLLMs-v3-7B-Chat 0.09 0.07 13.14 -0.28
Vistral-7B-Chat - - 167.07 -195.89
Few-shot
Qwen2.5-7B-Instruct 0.47 0.42 11.45 0.03
Llama-3.1-8B-Instruct 0.42 0.36 14.73 -0.60
Ministral-8B-Instruct-2410 0.32 0.20 46.76 -15.26
Llama-SEA-LION-v3-8B-IT 0.45 0.38 10.65 0.16
Sailor2-8B-Chat 0.44 0.42 11.25 0.07
SeaLLMs-v3-7B-Chat 0.40 0.34 20.80 -2.19
Finetuned
Qwen2.5-7B-Instruct 0.34 0.33 11.47 0.03
Llama-3.1-8B-Instruct 0.30 0.32 10.62 0.17
Ministral-8B-Instruct-2410 0.11 0.04 12.05 -0.07
Llama-SEA-LION-v3-8B-IT 0.30 0.33 10.60 0.17
Sailor2-8B-Chat 0.30 0.30 13.31 -0.30
SeaLLMs-v3-7B-Chat 0.32 0.31 10.75 0.15
Table 3: Performance comparison of various large language models (LLMs) across 2 tasks: Job Classifi-
cation and Salary Estimation
7. Conclusion and Future Work
This paper introduced VietJobs, a large-scale dataset of
48,092 Vietnamese job advertisements for research in
job classification, salary estimation, and labour market
analysis. The dataset covers 16 normalised job cat-
egories and includes structured fields such as job ti-
tles, salaries, skills, and employment conditions. Its util-
ity was evaluated through two core tasks using large
language models, with instruction-tuned systems such
as Qwen2.5-7B-Instruct and Llama-SEA-LION-v3-8B-IT
achieving the most consistent results. Fine-tuning on
combined datasets produced the most accurate salary
estimates.
VietJobs provides

yment conditions. Its util-
ity was evaluated through two core tasks using large
language models, with instruction-tuned systems such
as Qwen2.5-7B-Instruct and Llama-SEA-LION-v3-8B-IT
achieving the most consistent results. Fine-tuning on
combined datasets produced the most accurate salary
estimates.
VietJobs provides a foundation for future research in
computational labour market monitoring and recruitment
modelling. Future work may expand coverage across
additional platforms and time periods, incorporate mul-
tilingual or demographic data, and explore advanced
modelling methods such as retrieval-augmented gener-
ation or domain-adaptive pretraining. In addition, while
this study focused exclusively on the performance of
large language models, future research should conduct
systematic comparisons with traditional machine learn-
ing approaches, such as TF-IDF feature representations
combined with classifiers like Logistic Regression. Such
comparisons would help quantify the relative benefits
of LLM-based methods over conventional baselines in
Vietnamese job classification and salary prediction tasks.
Ethical Considerations and Limitations
All data used in this study were collected from publicly
accessible pages of the TopCV.vn website, which does
not prohibit web scraping under its robots.txt policy.
Only non-personal, publicly available job advertisement
content was included; no user profiles, résumés, or pri-
vate materials were accessed. The dataset contains no
personally identifiable information (PII), and all text was
processed solely for research purposes in accordance
with fair-use and data protection principles. This study
received ethical approval from an Institutional Ethical
Review Board under the principal investigator: Mo El-
Haj. The approval (Decision No. J/2025/CN/HDDD) was
granted in accordance with Circular 43/2024/TT-BYT,
which regulates the establishment and operation of ethics
councils in research, alongside relevant institutional poli-
cies.

proval from an Institutional Ethical
Review Board under the principal investigator: Mo El-
Haj. The approval (Decision No. J/2025/CN/HDDD) was
granted in accordance with Circular 43/2024/TT-BYT,
which regulates the establishment and operation of ethics
councils in research, alongside relevant institutional poli-
cies. The study was classified as minimal risk and au-
thorised for the period October–December 2025. All re-
search activities adhered to institutional guidelines, Viet-

-- 7 of 10 --

LLM VietJobs Vietnam Jobs Dataset Both
RMSE R2 RMSE R2 RMSE R2
Base Model
Qwen2.5-7B-Instruct 14.06 -0.46 16.19 -0.31 15.61 -0.34
Llama-3.1-8B-Instruct 16.01 -0.87 1319.94 -7744.05 1092.21 -6031.91
Ministral-8B-Instruct-2410 40.75 -13.12 2600.09 -30621.95 2190.20 -25135.32
Llama-SEA-LION-v3-8B-IT 11.72 0.07 15.78 -0.05 14.61 -0.02
Sailor2-8B-Chat 13.74 -3.29 99.92 -12.80 85.19 -12.55
SeaLLMs-v3-7B-Chat 13.14 -0.28 17.93 -0.08 15.82 -0.14
Few-shot
Qwen2.5-7B-Instruct 11.45 0.03 15.43 -0.18 14.40 -0.14
Llama-3.1-8B-Instruct 14.73 -0.60 20.80 -1.16 19.25 -1.04
Ministral-8B-Instruct-2410 46.76 -15.26 45.66 -9.48 45.99 -10.72
Llama-SEA-LION-v3-8B-IT 10.65 0.16 14.57 -0.06 13.40 0.01
Sailor2-8B-Chat 11.25 0.07 15.62 -0.21 14.53 -0.15
SeaLLMs-v3-7B-Chat 20.80 -2.19 21.78 -1.37 21.50 -1.54
Fine-tuned on VietJobs
Qwen2.5-7B-Instruct 11.47 0.03 14.28 -0.02 13.48 0
Llama-3.1-8B-Instruct 10.62 0.17 13.93 0.03 13.08 0.06
Ministral-8B-Instruct-2410 12.05 -0.07 14.62 -0.07 13.94 -0.07
Llama-SEA-LION-v3-8B-IT 10.60 0.17 13.89 0.04 13.04 0.07
Sailor2-8B-Chat 13.31 -0.30 16.84 -0.41 15.97 -0.40
SeaLLMs-v3-7B-Chat 10.75 0.15 14.05 0.02 13.20 0.04
Fine-tuned on Vietnam Jobs Dataset
Qwen2.5-7B-Instruct 10.77 0.15 13.40 0.10 12.70 0.11
Llama-3.1-8B-Instruct 10.82 0.14 13.34 0.11 12.68 0.12
Ministral-8B-Instruct-2410 11.70 -0.01 14.40 -0.03 13.69 -0.03
Llama-SEA-LION-v3-8B-IT 10.70 0.16 13.24 0.12 12.58 0.13
Sailor2-8B-Chat 10.94 0.12 13.78 0.05 13.03 0.07
SeaLLMs-v3-7B-Chat 10.83 0.14 13.44 0.10

t
Qwen2.5-7B-Instruct 10.77 0.15 13.40 0.10 12.70 0.11
Llama-3.1-8B-Instruct 10.82 0.14 13.34 0.11 12.68 0.12
Ministral-8B-Instruct-2410 11.70 -0.01 14.40 -0.03 13.69 -0.03
Llama-SEA-LION-v3-8B-IT 10.70 0.16 13.24 0.12 12.58 0.13
Sailor2-8B-Chat 10.94 0.12 13.78 0.05 13.03 0.07
SeaLLMs-v3-7B-Chat 10.83 0.14 13.44 0.10 12.75 0.11
Fine-tuned on Both
Qwen2.5-7B-Instruct 11.32 0.06 13.45 0.10 12.88 0.09
Llama-3.1-8B-Instruct 10.31 0.22 13.20 0.13 12.44 0.15
Ministral-8B-Instruct-2410 11.85 -0.03 14.29 -0.02 13.64 -0.02
Llama-SEA-LION-v3-8B-IT 10.24 0.23 13.17 0.13 12.40 0.16
Sailor2-8B-Chat 12.13 -0.08 13.73 0.06 13.34 0.02
SeaLLMs-v3-7B-Chat 10.48 0.19 13.41 0.10 12.64 0.12
Table 4: Performance comparison of LLMs in Salary Estimation with different settings
namese national regulations, and internationally recog-
nised ethical standards, including the principles of the
Declaration of Helsinki. No personal or sensitive data
were collected, and no human subjects were involved.
References
Abdel Rahman Alsheyab, Mohammad Alkhasawneh, and
Nidal Shahin. 2025. Job market cheat codes: Proto-
typing salary prediction and job grouping with synthetic
job listings.
Sarah H Bana. 2022. Work2vec: using language models
to understand wage premia.
BlossomsAI. 2025. Bloomvn-8b-chat.
Giuliano Bonoli and Karl Hinrichs. 2012. Statistical dis-
crimination and employers’recruitment: Practices for
low-skilled workers. European Societies, 14(3):338–
361.
Adwin Chan, Nicholas Cheng, Esther Choa, Yuli Huang,
Adithya Venkatadri Hulagadri, Wayne Lau, Chwan Ren
Lee, Wai Yi Leong, Wei Qi Leong, Peerat Limkon-
chotiwat, Bing Jie Darius Liu, Jann Railey Monta-
lan, Boon Cheong Raymond Ng, Jian Gang Ngui,
Thanh Ngan Nguyen, Brandon Ong, Tat-Wee David
Ong, Zhi Hao Ong, Hamsawardhini Rengarajan,
Bryan Siow, Yosephine Susanto, Ngee Chia Tai,
Choon Meng Tan, Walter Teng, Eng Sipp Leslie

-- 8 of 10 --

Teo, Wei Yi Teo, William Tjhi, Yeow Tong Yeo,
and Xianbin Yong. 2024. Llama-SEA-LION-v3-
8B-IT: Southeast

Jian Gang Ngui,
Thanh Ngan Nguyen, Brandon Ong, Tat-Wee David
Ong, Zhi Hao Ong, Hamsawardhini Rengarajan,
Bryan Siow, Yosephine Susanto, Ngee Chia Tai,
Choon Meng Tan, Walter Teng, Eng Sipp Leslie

-- 8 of 10 --

Teo, Wei Yi Teo, William Tjhi, Yeow Tong Yeo,
and Xianbin Yong. 2024. Llama-SEA-LION-v3-
8B-IT: Southeast asian llm with instruction tun-
ing. https://huggingface.co/aisingapore/
Llama-SEA-LION-v3-8B-IT.
Gheorghe Comanici, Eric Bieber, Mike Schaekermann,
Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Mar-
cel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al.
2025. Gemini 2.5: Pushing the frontier with ad-
vanced reasoning, multimodality, long context, and
next generation agentic capabilities. arXiv preprint
arXiv:2507.06261.
Nikolas Dawson, Marian-Andrei Rizoiu, Benjamin John-
ston, and Mary-Anne Williams. 2020. Predicting skill
shortages in labor markets: A machine learning ap-
proach. In 2020 IEEE International Conference on Big
Data (Big Data), pages 3052–3061. IEEE.
Longxu Dou, Qian Liu, Fan Zhou, Changyu Chen, Zili
Wang, Ziqi Jin, Zichen Liu, Tongyao Zhu, Cunxiao Du,
Penghui Yang, et al. 2025. Sailor2: Sailing in south-
east asia with inclusive multilingual llms. arXiv preprint
arXiv:2502.12982.
Nazarii Drushchak and Mariana Romanyshyn. 2024. In-
troducing the djinni recruitment dataset: A corpus of
anonymized CVs and job postings. In Proceedings
of the Third Ukrainian Natural Language Processing
Workshop (UNLP) @ LREC-COLING 2024, pages
8–13, Torino, Italia. ELRA and ICCL.
Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey,
Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman,
Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan,
et al. 2024. The llama 3 herd of models. arXiv e-prints,
pages arXiv–2407.
Mo El-Haj. 2025. Arabjobs: A multinational corpus of
arabic job ads. In Proceedings of the Third Arabic
Natural Language Processing Conference (ArabicNLP
2025), Suzhou, China. Association for Computational
Linguistics. Co-located with EMNLP 2025.
Aaron Hurst,

e llama 3 herd of models. arXiv e-prints,
pages arXiv–2407.
Mo El-Haj. 2025. Arabjobs: A multinational corpus of
arabic job ads. In Proceedings of the Third Arabic
Natural Language Processing Conference (ArabicNLP
2025), Suzhou, China. Association for Computational
Linguistics. Co-located with EMNLP 2025.
Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perel-
man, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila
Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-
4o system card.
Tanjila Kanij, John Grundy, and Jennifer McIntosh. 2024.
Enhancing understanding and addressing gender bias
in it/se job advertisements. Journal of Systems and
Software, 217:112169.
Kanchana Karunarathna. 2025. Adzuna global job list-
ings 2025.
Jackson G Lu. 2023. Asians don’t ask? relational con-
cerns, negotiation propensity, and starting salaries.
Journal of Applied Psychology, 108(2):273.
Antoine Magron, Anna Dai, Mike Zhang, Syrielle Montar-
iol, and Antoine Bosselut. 2024. JobSkape: A frame-
work for generating synthetic job postings to enhance
skill matching. In Proceedings of the First Workshop on
Natural Language Processing for Human Resources
(NLP4HR 2024), pages 43–58, St. Julian’s, Malta.
Association for Computational Linguistics.
Harry Moroz and Nga Thi Nguyen. 2019. Skills profiling
of priority occupations in vietnam. Technical report,
World Bank.
Chitinh Nguyen. 2025. Vietnam jobs dataset.
Dat Quoc Nguyen, Linh The Nguyen, Chi Tran,
Dung Ngoc Nguyen, Dinh Phung, and Hung Bui.
2023. PhoGPT: Generative Pre-training for Viet-
namese. arXiv preprint, arXiv:2311.02945.
Naoki Otani, Nikita Bhutani, and Estevam Hruschka.
2024. Natural language processing for human re-
sources: A survey. arXiv preprint arXiv:2410.16498.
Le Anh Tu Packard. 2006. Gender dimensions of Viet
Nam’s comprehensive macroeconomic and structural
reform policies. 14. UNRISD Occasional Paper.
Carlo Perroni, Kimberly Scharf, Oleksandr Talavera, and
Linh Vi. 2023. Gender beauty premia in wage offers:
Evidence from

sources: A survey. arXiv preprint arXiv:2410.16498.
Le Anh Tu Packard. 2006. Gender dimensions of Viet
Nam’s comprehensive macroeconomic and structural
reform policies. 14. UNRISD Occasional Paper.
Carlo Perroni, Kimberly Scharf, Oleksandr Talavera, and
Linh Vi. 2023. Gender beauty premia in wage offers:
Evidence from vietnamese online job postings.
Pieterbas Pluijmaekers and Francesco Lelli. 2022. A
dataset containing job descriptions suitable for nlp
and nn processing.
Qwen, :, An Yang, Baosong Yang, Beichen Zhang,
Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li,
Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian
Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi
Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming
Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng
Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tian-
hao Li, Tianyi Tang, Tingyu Xia, Xingzhang Ren, Xu-
ancheng Ren, Yang Fan, Yang Su, Yichang Zhang,
Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and
Zihan Qiu. 2025. Qwen2.5 technical report.
Rohan Prakash Sharma. 2025. Recognizing and ex-
plaining bias in job descriptions: A roberta-powered
recruitment framework. Science and Engineering Re-
search Journal, 13(3):1–9.
Granite Team. 2025. Granite-3.3-8b-instruct.
Viet-Trung Tran, Hai-Nam Cao, and Tuan-Dung Cao.
2022. A practical method for occupational skills detec-
tion in vietnamese job listings. In Asian Conference on
Intelligent Information and Database Systems, pages
571–581. Springer.
UncleCode. 2024. Crawl4ai: Open-source llm friendly
web crawler and scraper. https://github.com/
unclecode/crawl4ai.
Chien Van Nguyen, Thuat Nguyen, Quan Nguyen, Huy
Nguyen, Bj ¨orn Pl ¨uster, Nam Pham, Huu Nguyen,
Patrick Schramowski, and Thien Nguyen. 2023.
Vistral-7b-chat-towards a state-of-the-art large lan-
guage model for vietnamese.
N. Van Thang, J. M. Peiró, L. Q. Canh, V. González-
Romá, and V. Martínez-Tur. 2020. Vietnamese gradu-
ates’ labour market entry and employment: A tracer
study. Social Sciences,

m, Huu Nguyen,
Patrick Schramowski, and Thien Nguyen. 2023.
Vistral-7b-chat-towards a state-of-the-art large lan-
guage model for vietnamese.
N. Van Thang, J. M. Peiró, L. Q. Canh, V. González-
Romá, and V. Martínez-Tur. 2020. Vietnamese gradu-
ates’ labour market entry and employment: A tracer
study. Social Sciences, 12(2):94.

-- 9 of 10 --

Jan Vogt, Thilo Voigt, Annika Nowak, and Jan M
Pawlowski. 2023. Development of a job advertise-
ment analysis for assessing data science competen-
cies. Data Science Journal, 22(1).
Dinh-Hong Vu, Kien Nguyen, Khai Thien Tran, Bay Vo,
and Tuong Le. 2025. Improving fake job descrip-
tion detection using deep learning-based nlp tech-
niques. Journal of Information and Telecommunica-
tion, 9(1):113–125.
Mike Zhang, Kristian Nørgaard Jensen, Sif Dam Sonniks,
and Barbara Plank. 2022. Skillspan: Hard and soft
skill extraction from english job postings. arXiv preprint
arXiv:2204.12811.
Wenxuan Zhang, Hou Pong Chan, Yiran Zhao, Ma-
hani Aljunied, Jianyu Wang, Chaoqun Liu, Yue Deng,
Zhiqiang Hu, Weiwen Xu, Yew Ken Chia, et al. 2024.
Seallms 3: Open foundation and chat multilingual
large language models for southeast asian languages.
arXiv preprint arXiv:2407.19672.

-- 10 of 10 --