DentalCorpus: Bilingual Dental Text Corpus

• DentalCorpus is a curated, bilingual corpus for dental language processing that supports domain adaptation for large language models.
• Its robust filtering methodology uses TF-IDF, glossary intersection, and human review to ensure high domain relevance and quality.
• The resource underpins both supervised fine-tuning and retrieval-augmented generation, delivering measurable gains in dental QA benchmarks.

DentalCorpus is a large-scale, high-quality corpus dedicated to the dental domain and introduced as one of the two principal components of DentalBench, alongside DentalQA. It was designed to support domain adaptation for LLMs in dentistry, with explicit support for supervised fine-tuning and retrieval-augmented generation across English and Chinese. In a literature where the dental AI data ecosystem has been dominated by imaging, mesh, and volumetric resources, DentalCorpus constitutes a textual infrastructure for dental language understanding, terminology handling, and knowledge-intensive question answering (Zhu et al., 28 Aug 2025).

1. Definition and scope

DentalCorpus is a professionally curated bilingual corpus for dental and oral healthcare language processing. Within DentalBench, it serves a different function from DentalQA: DentalQA is the evaluation benchmark with 36,597 questions spanning 4 tasks and 16 dental subfields, whereas DentalCorpus is the domain-adaptation resource used to improve model behavior in those dental tasks (Zhu et al., 28 Aug 2025).

Its stated purpose is twofold. First, it addresses the underexplored status of dentistry within LLM evaluation, despite strong progress on general medical benchmarks. Second, it provides an in-domain textual base for adaptation methods that require either retrieved passages at inference time or curated material for supervised task construction. The corpus is therefore not merely a passive document collection; it is a dataset engineered for downstream dental NLP workflows.

A common misunderstanding is to treat DentalCorpus as equivalent to DentalBench as a whole. The distinction is structurally important. DentalBench is the benchmark framework; DentalCorpus is the corpus component within that framework, paired with the bilingual QA benchmark rather than replacing it (Zhu et al., 28 Aug 2025).

2. Scale, language coverage, and internal organization

DentalCorpus contains 337.35 million tokens. Its final released composition is heavily English-dominant: 319.08M tokens in English and 18.27M tokens in Chinese. The material is segmented into chunks of up to 512 tokens each, yielding 1,060,000 English chunks and 66,300 Chinese chunks (Zhu et al., 28 Aug 2025).

This bilingual design is central to the corpus’s role in dental LLM evaluation. The benchmark explicitly targets English and Chinese, and the corpus mirrors that objective. At the same time, the token counts show that the bilinguality is asymmetrical rather than balanced. That asymmetry is a factual property of the resource rather than an incidental detail, because it conditions retrieval coverage, adaptation density, and terminology exposure across the two languages.

The corpus was assembled from three principal source categories. The first is textbooks, including 40 dental textbooks and auxiliary materials in Chinese, with front matter, back matter, and non-content sections removed; this source also includes a bilingual glossary of 1,971 dental terms. The second is PubMed full-text articles, obtained through 28 dentistry-related MeSH terms, with 54,651 full-text articles retrieved from 2000–2024. The third is open medical datasets, specifically MMedC and MedRAG (Zhu et al., 28 Aug 2025).

Because the corpus is chunked rather than preserved only as long-form documents, it is directly usable for dense retrieval and prompt construction. The segmentation policy also makes the collection compatible with vector indexing and passage-level search, which is the operative mode used in the reported RAG pipeline.

3. Source acquisition and curation methodology

The curation procedure is explicitly dental-specific. It begins with manual and professional source selection, but the decisive stage is a filtering pipeline meant to exclude general medical “bleed-through” and retain material with high dental relevance. That pipeline is based on TF-IDF analysis conducted on dental and general medical corpora, followed by construction of a candidate vocabulary from PubMed, MedRAG, and textbooks, intersected with the dental glossary (Zhu et al., 28 Aug 2025).

Terms that were overly frequent in non-dental general medical texts were excluded. The resulting keyword inventories consisted of 440 English and 235 Chinese terms. Chunks then had to satisfy two explicit criteria: more than 1% keyword density, and at least 2 distinct keywords per sentence. Tokenization was language-specific: English was tokenized by spaces, while Chinese used matched substrings (Zhu et al., 28 Aug 2025).

The post-filtering pipeline included MD5 hash deduplication, semantic processing with bge-m3 embeddings, and text segmentation to the 512-token maximum. Quality control included human review of 100 random samples in each language. The reported domain relevance was 99% for English and 96% for Chinese (Zhu et al., 28 Aug 2025).

These procedural details are significant because they distinguish DentalCorpus from a raw web crawl or a generic medical corpus subset. The resource is curated not only by topic selection but by explicit lexical density thresholds, glossary intersection, deduplication, and human validation. In that sense, DentalCorpus encodes a strong editorial stance about what constitutes dental-domain evidence.

4. Role in LLM adaptation workflows

DentalCorpus is used in three adaptation settings: SFT, RAG, and SFT + RAG. In the reported experiments with Qwen-2.5-3B-Instruct, supervised fine-tuning was run on the QA split for 4 epochs with learning rate 1e-6, batch size 16, and bfloat16 precision. The RAG pipeline retrieved the top-5 most relevant passages from DentalCorpus using FAISS + bge-m3 embeddings, and prepended those passages to the input prompt at inference time (Zhu et al., 28 Aug 2025).

The paper characterizes the corpus as supporting both supervised fine-tuning and retrieval-augmented generation. In the reported setup, this support is realized in complementary ways. Retrieval uses DentalCorpus directly as the inference-time knowledge store. Supervised adaptation uses labeled data from DentalQA, which the study describes as being built using DentalCorpus resources. The combined setting first applies SFT and then uses RAG during inference (Zhu et al., 28 Aug 2025).

Functionally, DentalCorpus supplies in-domain knowledge under two different operational logics. Under RAG, it acts as an external memory that can inject dental terminology and factual context at query time. Under the broader adaptation regime, it also underwrites the construction and improvement of dental task data. This dual role is why the corpus is framed as a domain-adaptation resource rather than only a retrieval database.

The reported results also indicate that DentalCorpus is particularly consequential for knowledge-intensive and terminology-focused tasks. That point is important because dentistry is not only a subspecialty of medicine; it is also a domain with dense nomenclature, procedural vocabulary, and cross-lingual terminological variation, all of which make corpus quality more important than raw parameter scale alone (Zhu et al., 28 Aug 2025).

5. Empirical impact on benchmark performance

The empirical case for DentalCorpus rests on adaptation experiments conducted on DentalQA. The study evaluated 14 LLMs, covering proprietary, open-source, and medical-specific models, and reported significant performance gaps across task types and languages. The experiments with Qwen-2.5-3B-Instruct provide the clearest quantitative demonstration of corpus-enabled gains (Zhu et al., 28 Aug 2025).

On MCQ (ACC), the baseline model scored 48.63 in Chinese and 41.77 in English. With +SFT, these rose to 54.58 and 47.90. With +RAG, they reached 54.45 and 48.74. In the SFT + RAG configuration, MCQ accuracy rose to 60.06 in Chinese and 52.15 in English (Zhu et al., 28 Aug 2025).

On OEQ (BERTScore), the baseline values were 20.89 in Chinese and 34.48 in English. The +SFT setting produced 22.42 and 37.74, while +RAG produced 30.18 and 36.47. The combined setting reported 30.06 in Chinese and 37.68 in English. On DEF (BERTScore), the baseline values were 11.16 in Chinese and 29.62 in English; +SFT yielded 15.29 and 30.79, +RAG yielded 22.13 and 30.04, and SFT + RAG yielded 20.85 and 29.65 (Zhu et al., 28 Aug 2025).

The study’s interpretation is that SFT and RAG both deliver improvements over vanilla LLMs, that RAG yields especially large benefits on open-ended and definition tasks, and that combined SFT+RAG is most effective, with additive improvements. It also reports that medical LLMs not fine-tuned on dental data perform worse than general LLMs after domain adaptation, underscoring the importance of dental-domain resources rather than assuming that general medical specialization is sufficient (Zhu et al., 28 Aug 2025).

These findings situate DentalCorpus as an enabling asset rather than a mere supplement. Its impact is measurable not only in multiple-choice accuracy but also in semantic fidelity on open-ended and definitional tasks, where lexical precision and domain-specific factual grounding are especially demanding.

6. Position within the broader dental data ecosystem

DentalCorpus emerged in a research landscape where major dental datasets have largely been visual or geometric. Examples include the 16,000 IOS dataset introduced for TSegFormer (Xiong et al., 2023), the 503 CBCT and 28,559 IOS meshes used in DDMA (Hao et al., 2022), the 950 upper jaw samples used for 3D oral reconstruction from 2D intraoral images (Cho et al., 4 Jun 2026), and the Teeth3DS benchmark used for dental point-cloud completion (Sun et al., 3 Dec 2025). Those resources support segmentation, reconstruction, fusion, and restoration; DentalCorpus instead targets dental language adaptation.

This suggests that DentalCorpus occupies the textual layer of a wider dental AI stack. In that stack, imaging and mesh corpora support perception and geometry, while DentalCorpus supports terminology, explanation, question answering, and retrieval-grounded inference. The distinction is methodologically important because clinical dental systems increasingly combine structured geometry with language interfaces, decision support, or documentation workflows.

Its significance is therefore not limited to benchmark construction. The paper explicitly presents the resource as necessary for developing trustworthy and effective LLMs tailored to healthcare applications. That claim is consistent with the empirical finding that adaptation materially improves dental performance, especially on domain-specific tasks (Zhu et al., 28 Aug 2025).

At the same time, the corpus has visible boundaries. It is bilingual but not balanced, with much greater English volume than Chinese. It is dental-focused but built partly from broader medical sources that had to be heavily filtered. It supports SFT and RAG, but its reported use is embedded in a benchmark-and-adaptation framework rather than a fully open-ended clinical deployment study. These properties do not diminish its role; they define the specific problem it was built to solve.

In sum, DentalCorpus is a curated bilingual dental text resource designed to make dentistry legible to LLMs at benchmark, adaptation, and retrieval time. Its scale, filtering methodology, and measured downstream gains position it as a foundational corpus for dental-domain NLP, and as a textual counterpart to the imaging and geometric datasets that have previously defined digital dentistry research (Zhu et al., 28 Aug 2025).