Vietnamese NLP Datasets: What Exists, What's Missing, and Where to Get Them

Vietnamese is the 14th most spoken language in the world with approximately 95 million native speakers, yet it remains severely underrepresented in the training data for large language models and production NLP systems. English-dominant foundation models (GPT-4, Claude, Llama) have limited Vietnamese capability compared to their English performance, and the gap is most pronounced for tasks requiring cultural knowledge, colloquial language, and domain-specific vocabulary.

The core problem is training data asymmetry. English NLP benefits from Common Crawl, BooksCorpus, Wikipedia, and decades of commercial annotation investment. Vietnamese has a growing but still sparse ecosystem of public datasets, most of which were created for academic benchmarks rather than production NLP pipelines.

This guide covers the publicly available Vietnamese NLP datasets, their limitations, and the options for teams that need production-quality Vietnamese training data for LLM fine-tuning, conversational AI, and domain-specific NLP applications.

The most widely used public Vietnamese NLP resources include PhoNLP (part-of-speech tagging and NER), VLSP (Vietnamese Language and Speech Processing consortium benchmarks), UIT-VSFC (sentiment analysis of student feedback), UIT-ViCTSD (constructive and toxic speech detection), and PhoMT (machine translation parallel corpus).

PhoNLP and the VLSP benchmark datasets are essential references for Vietnamese NLP research, but they are too narrow for production LLM fine-tuning. They cover standard benchmark tasks on formal Vietnamese text, not the colloquial, mixed-language, and domain-specific Vietnamese that appears in real commercial applications.

Oscar Corpus includes Vietnamese text from web crawls, but as with all web-crawled corpora, Vietnamese quality is variable - the web crawl methodology that works for English does not transfer well to Vietnamese, where code-switching (mixing Vietnamese and English), diacritical markup issues, and spam content create significant noise in the raw corpus.

Production Vietnamese NLP applications consistently encounter four data gaps that public datasets cannot fill. First, conversational Vietnamese: the spoken and informal written registers of Vietnamese differ substantially from formal written Vietnamese, and most public datasets are drawn from formal sources (news, Wikipedia, academic text). Chatbots and voice assistants trained on formal Vietnamese produce responses that sound stilted to native speakers.

Second, domain-specific vocabulary: Vietnamese finance, legal, healthcare, and technology vocabulary has evolved rapidly over the past decade with substantial English loanword adoption, domain-specific abbreviations, and regulatory terminology. Generic Vietnamese corpora do not adequately represent these domain registers.

Third, Southern and Northern dialect variation: formal written Vietnamese is largely standardized, but spoken Vietnamese and informal written Vietnamese (SMS, social media, customer service chats) differ substantially between Northern and Southern registers. Models trained on standard Vietnamese perform poorly on Southern dialect content.

Fourth, code-switching: real Vietnamese digital communication heavily mixes Vietnamese and English, particularly in technology, business, and youth demographics. "Em đã check API và thấy error 403 rồi anh" is a typical message in a Vietnamese tech support context. Models trained on pure-Vietnamese or pure-English corpora handle this poorly.

Commercial Vietnamese NLP datasets from vendors like DataX Power address the gaps that academic datasets leave open. Key differentiators in commercial collections include: (1) conversational register coverage - datasets collected from customer service dialogues, chat transcripts (with consent), and spoken conversation transcription; (2) domain-specific vocabulary coverage for finance, healthcare, legal, and e-commerce domains; (3) dialect labeling that identifies Northern vs. Southern vs. Central register; (4) code-switching annotation for text that mixes Vietnamese and English.

For LLM fine-tuning specifically, Vietnamese instruction-tuning datasets need to match the task types, response style, and formatting conventions that make Vietnamese responses feel natural to native speakers - not just grammatically correct translated English. This requires native speaker design of instruction-response pairs, not translation of English SFT datasets.

Vietnamese preference data for RLHF is the most constrained commercial category. Native speaker preference evaluators for Vietnamese RLHF require both language competence and the cultural knowledge to evaluate whether AI responses are appropriate for Vietnamese cultural contexts - not just whether they are grammatically correct.

The data gaps for other major Southeast Asian languages follow similar patterns to Vietnamese but with language-specific nuances. Thai NLP has stronger institutional support (National Electronics and Computer Technology Center, several Thai university NLP labs) and larger public benchmark datasets than Vietnamese, but commercial domain-specific datasets are similarly sparse.

Indonesian (Bahasa Indonesia) benefits from its status as a high-resource language in multilingual models due to web crawl volume, but colloquial Indonesian (Bahasa Gaul), regional dialect mixing (Javanese, Sundanese, Betawi influence in Jakarta), and domain-specific Indonesian NLP data gaps are significant for production applications.

Malay (Bahasa Melayu) has overlap with Indonesian at the formal register level but diverges substantially in colloquial speech, government and legal terminology, and cultural reference. Models fine-tuned on Indonesian data perform poorly on Malay content and vice versa, particularly in formal and institutional domains.

For teams that need Vietnamese NLP training data that public and commercial pre-built datasets cannot provide, custom collection and annotation is the path. The typical program structure for Vietnamese LLM fine-tuning data: (1) Define the task types - instruction following, question answering, summarization, sentiment analysis, NER, or dialogue; (2) Define the register - formal, conversational, or mixed; (3) Define the domain - general, finance, healthcare, legal, or e-commerce; (4) Define the dialect requirements - standardized, Northern, Southern, or all three.

Text collection sources for Vietnamese corpora include licensed news archive partnerships, customer service transcript collections (with enterprise consent frameworks), and native speaker writing programs where annotators produce original text in specified registers and domains. Web scraping without careful Vietnamese-specific filtering produces too much noise to be cost-effective for annotation.

Native speaker annotators for Vietnamese NLP programs must be recruited with language competence testing that goes beyond simple literacy. For sentiment, intent, and cultural appropriateness tasks, cultural competence matters as much as language ability. Programs using annotators from one regional background to annotate content from another region consistently produce lower IAA scores and less reliable labels.

DataX Power runs Vietnamese NLP annotation programs from Hanoi with annotator networks across Northern and Southern Vietnam, covering both formal and colloquial registers and major domain categories including finance, healthcare, e-commerce, and technology.