GLiNER-BioMed: Open Biomedical NER Models

• GLiNER-BioMed is a suite of efficient models for open biomedical NER that uses natural-language entity descriptions instead of fixed taxonomies.
• It integrates synthetic biomedical pretraining, domain adaptation, and dual encoder architectures (uni-encoder and bi-encoder) to achieve a 5.96% absolute F1-score improvement over strong baselines.
• The approach enables scalable, zero-shot and few-shot recognition across diverse biomedical datasets, supporting rapid adaptation to emerging and specialized entity classes.

GLiNER-BioMed is a suite of efficient models for open biomedical named entity recognition that adapts the GLiNER framework to specialized biomedical text. It is designed for zero-shot and few-shot recognition of arbitrary biomedical entity types through natural-language labels, rather than through a fixed taxonomy. The system combines domain adaptation, synthetic biomedical NER pretraining distilled from large biomedical LLMs, and two encoder architectures—uni-encoder and bi-encoder—to address specialized vocabularies, heterogeneous biomedical genres, and continuously emerging entity classes. Across several biomedical datasets, it reports a 5.96% absolute improvement in F1-score over the strongest baseline, with a reported $p < 0.001$ (Yazdani et al., 1 Apr 2025).

1. Conceptual basis and problem setting

GLiNER-BioMed addresses a biomedical NER regime in which entity inventories are large, dynamic, and often task-specific. The motivating challenges are explicit: biomedical text contains specialized and dense vocabulary, ontologies such as UMLS contain millions of concepts, and biomedical language varies substantially across drug labels, clinical notes, patents, clinical trials, social media, and PubMed abstracts (Yazdani et al., 1 Apr 2025). In that setting, closed-set models parameterized over a fixed label inventory are poorly matched to the continual introduction of new drugs, genes, variants, devices, and task-defined concepts.

The model inherits its basic orientation from GLiNER, which introduced a compact NER architecture trained to identify any type of entity with a bidirectional transformer encoder and natural-language entity descriptions (Zaratiana et al., 2023). In the original GLiNER formulation, NER is not tied to predefined labels such as [PER](https://www.emergentmind.com/topics/preference-informed-edge-reweighting-per) or ORG; instead, arbitrary entity types are provided as textual prompts, and the model performs semantic matching between text spans and label descriptions. GLiNER was already competitive in zero-shot biomedical settings, reporting, for example, 47.9 on BC2GM, 66.4 on BC5CDR, 55.5 on GENIA, and 61.9 on NCBI Disease (Zaratiana et al., 2023). GLiNER-BioMed should therefore be understood not as a departure from GLiNER, but as a biomedical specialization of an already open-type NER paradigm.

Within the biomedical paper, the target setting is explicitly “open” or “arbitrary-label” NER: the system should identify entities whose types are described only in natural language at inference time. The example given is “Find all mentions of ‘non-pharmacological interventions’ such as physical therapy or cognitive behavioral therapy” (Yazdani et al., 1 Apr 2025). This suggests a shift from ontology-bound tagging toward task-conditioned extraction, where label semantics are part of the input rather than hard-coded into a classifier head.

2. Model family and architectural variants

GLiNER-BioMed comprises two architectural families. The uni-encoder variant retains the original GLiNER pattern of concatenating text and label descriptions into a single sequence and encoding them jointly with a DeBERTa-v3 backbone at small, base, and large scales. In this configuration, token–label interactions are mediated by shared self-attention, and computational complexity is given as $\mathcal{O}((n_t + n_e)^2)$, where $n_t$ is the number of text tokens and $n_e$ the number of label tokens (Yazdani et al., 1 Apr 2025).

The bi-encoder variant, GLiNER-BioMed-bi, separates text and label encoding. Its text encoder is DeBERTa-v3 at small, base, or large scale, while its label encoder is an independent sentence-embedding model: all-MiniLM-L6-v2 for the small model, bge-small-en-v1.5 for the base model, and bge-base-en-v1.5 for the large model (Yazdani et al., 1 Apr 2025). In this design, label embeddings can be pre-computed and cached, and complexity becomes $\mathcal{O}(n_t^2 + n_e)$, which is materially more scalable for large label inventories.

The biomedical paper presents the scoring mechanism as similarity between text and label representations, using

$s_{t,i} = h_t^\top z_i$

with prediction derived from a thresholded classifier output (Yazdani et al., 1 Apr 2025). More generally, the architecture retains the GLiNER principle of representation matching between textual content and natural-language label descriptions rather than softmax classification over a fixed schema. A practical implication stated in the paper is that the bi-encoder is more suitable for large ontologies, including “thousands of UMLS concepts,” while the uni-encoder remains simpler and, at large scale, stronger in overall performance (Yazdani et al., 1 Apr 2025).

The division of labor between the two families is empirical rather than purely theoretical. At small and base scales, the bi-encoder outperforms the uni-encoder; at large scale, the uni-encoder performs best (Yazdani et al., 1 Apr 2025). The paper hypothesizes that separate encoders effectively increase capacity at small and base sizes, whereas large uni-encoders can better exploit cross-attention between labels and text.

3. Synthetic biomedical data generation and training pipeline

A defining feature of GLiNER-BioMed is its synthetic biomedical pretraining corpus. The authors curate approximately 115,000 passages from five sources after filtering and deduplication: human prescription labels from DailyMed, clinical trial descriptions and arm-level regimens from ClinicalTrials.gov, PubMed abstracts under the MeSH term “pathological conditions, signs, and symptoms,” and patents from WIPO IPC categories A61P, A61K, and G16H (Yazdani et al., 1 Apr 2025).

The corpus construction pipeline includes heuristic quality filtering and graph-based deduplication. Filtering uses criteria such as special character frequency, number of sentences, average words per sentence, uppercase/lowercase ratio, lexical diversity, stopword prevalence, repetition, and newline-to-sentence ratio. Deduplication represents passages with TF-IDF vectors, constructs a similarity graph with cosine similarity greater than 0.9, and retains only the most central passage in each connected component (Yazdani et al., 1 Apr 2025). The intent is to reduce redundancy while preserving source diversity.

Annotation is produced by distillation from large biomedical LLMs. First, 10,000 sampled passages are annotated by OpenBioLLM-70B in a few-shot conversational setup. The model is instructed to annotate all noun phrases with entity type(s), rather than performing unconstrained free-span extraction, and its JSON output is constrained by logit processing (Yazdani et al., 1 Apr 2025). Second, OpenBioLLM-8B is fine-tuned with LoRA on those 10,000 annotations to reproduce the teacher’s behavior. Third, the distilled 8B model annotates the remaining roughly 105,000 passages. The resulting synthetic pretraining set contains 2.3 million entity mentions and 640,000 unique entities (Yazdani et al., 1 Apr 2025).

GLiNER-BioMed does not stop at synthetic biomedical pretraining. It also uses a post-training stage on a 19,000-instance mixed dataset composed of a synthetic FineWeb subset annotated by Qwen2.5-72B and human-annotated OntoNotes 5.0, MultiNERD, and WNUT2017 (Yazdani et al., 1 Apr 2025). Biomedical evaluation entity types are removed from this post-training set to avoid leakage into zero-shot evaluation. This two-stage regime is central to the model’s behavior: synthetic biomedical pretraining supplies broad domain coverage, while post-training improves label handling, boundary quality, and general zero-shot robustness.

Ablation results make this interaction explicit. Synthetic biomedical data alone yields high precision and low recall—Precision 70.08, Recall 30.09, F1 42.10—whereas synthetic biomedical pretraining followed by post-training yields Precision 56.67, Recall 63.22, F1 59.77 (Yazdani et al., 1 Apr 2025). The paper’s interpretation is that synthetic pretraining injects biomedical knowledge but is too conservative when used in isolation.

4. Evaluation benchmark and empirical results

The evaluation benchmark spans eight datasets, totaling 10,918 samples, 85,959 mentions, and 58 unique entity types: TAC 2017 ADE, CADEC, N2C2 2018, BC5CDR, BioRED, CHIA, Biomed NER, and NCBI Disease (Yazdani et al., 1 Apr 2025). These datasets cover regulatory documents, informal health forums, clinical narratives, clinical trial eligibility criteria, and research abstracts. Evaluation uses exact span match at span level, with micro F1, macro mean F1, macro median F1, and weighted F1 (Yazdani et al., 1 Apr 2025).

The strongest zero-shot result is obtained by GLiNER-biomed-large, the large uni-encoder. Against the best general-domain baseline, GLiNER v2.5-large, it improves micro F1 from 53.81 to 59.77, macro mean F1 from 35.22 to 40.67, macro median F1 from 35.65 to 42.65, and weighted F1 from 51.57 to 58.40 (Yazdani et al., 1 Apr 2025).

Model	Micro F1	Note
GLiNER v2.5-large	53.81	Best general-domain baseline
GLiNER-biomed-large	59.77	Best large uni-encoder
GLiNER-biomed-bi-large	54.90	Large bi-encoder
GLiNER-biomed-base	54.37	Base uni-encoder
GLiNER-biomed-bi-base	58.31	Best base model
GLiNER-biomed-small	52.53	Small uni-encoder
GLiNER-biomed-bi-small	56.93	Best small model

The size-efficiency trade-off is notable. GLiNER-biomed-small, described as having 7× fewer parameters than GLiNER v2.5-large, achieves 52.53 micro F1 against 53.81 for that larger baseline (Yazdani et al., 1 Apr 2025). This is one of the clearest arguments in the paper that domain adaptation matters more than sheer scale.

Few-shot experiments reinforce that conclusion. Using 10, 20, and 50 samples per dataset, the best-performing model is GLiNER-biomed-bi-large, with micro F1 scores of 70.39, 73.07, and 76.02, respectively. Under full supervision, GLiNER-biomed-large reaches 84.95 and GLiNER-biomed-bi-large 84.91, while GLiNER v2.5-large reaches 84.64 (Yazdani et al., 1 Apr 2025). The pattern is consistent: biomedical adaptation helps most in zero-shot and few-shot regimes, while performance converges under full supervision.

5. Efficiency, deployment, and released resources

GLiNER-BioMed is explicitly positioned as an encoder-only alternative to large generative biomedical LLMs. Although the paper does not give exact parameter counts for each released model, it states that the DeBERTa-v3-based encoders are on the order of tens to hundreds of millions of parameters rather than billions, and that label encoders in the bi-encoder family are compact sentence-transformer models (Yazdani et al., 1 Apr 2025). Relative to 70B-scale annotation models, the shift in inference cost is substantial.

Efficiency depends on architecture. The uni-encoder is simpler but becomes expensive when label counts are high because labels share the same self-attention space as the text. The bi-encoder decouples those computations, allowing label embeddings to be cached and making the model suitable for large ontologies (Yazdani et al., 1 Apr 2025). The paper’s deployment guidance is correspondingly differentiated: GLiNER-biomed-large is recommended as the primary zero-shot model, while GLiNER-biomed-bi—especially at small and base scales—is recommended for few-shot or resource-constrained settings (Yazdani et al., 1 Apr 2025).

The release is unusually complete. The authors provide pretrained uni-encoder and bi-encoder models at small, base, and large scales; the full synthetic biomedical dataset with 2.3 million mentions and 640,000 entities; the 19,000-instance post-training dataset; and the synthetic data generation and training pipelines, including prompts and LoRA setups for LLM distillation, through the repository at https://github.com/ds4dh/GLiNER-biomed (Yazdani et al., 1 Apr 2025). This makes GLiNER-BioMed not only a model family but also a reproducible biomedical IE pipeline.

6. Limitations, later use, and interpretive significance

The main limitations identified in the GLiNER-BioMed paper concern annotation quality, coverage, and analysis depth. Synthetic labels are generated by a distilled 8B model and may therefore inherit biases and systematic errors from the 70B teacher. Certain biomedical subdomains, including veterinary medicine, dentistry, and rare specialties, remain under-represented in evaluation. The synthetic annotation and model training pipeline is compute-intensive despite efficient inference, and the paper emphasizes quantitative benchmarking more than qualitative failure analysis (Yazdani et al., 1 Apr 2025).

Subsequent use of GLiNER-BioMed in BioASQ 2025 Task 6 illustrates both the model’s strength as a biomedical NER backbone and the fragility of downstream heuristics. In that work, GLiNER-BioMed is fine-tuned on a 13-label GutBrainIE entity schema and used as the main NER engine. The baseline fine-tuned model reaches Dev micro-F1 0.7857, while targeted dictionary-based post-processing raises Dev micro-F1 to 0.8316; however, the same post-processed system achieves only 0.7743 micro-F1 on the blind test set (Mehta, 3 Oct 2025). The paper interprets this as overfitting to development-set error patterns, especially gene/chemical and food/dietary_supplement confusions.

That later result clarifies an important point about GLiNER-BioMed. The underlying model is strong enough to serve as a generalist biomedical recognizer in specialized shared-task settings, but symbolic refinements built on top of it can generalize poorly if they are tuned too narrowly to local error distributions (Mehta, 3 Oct 2025). A plausible implication is that GLiNER-BioMed is best understood as a robust open-label encoder backbone whose gains are most reliable when they derive from corpus-scale domain adaptation rather than from highly targeted post hoc rule systems.

Taken together, GLiNER-BioMed occupies a distinct position in biomedical information extraction. It retains the original GLiNER commitment to natural-language labels and efficient encoder-based matching (Zaratiana et al., 2023), but relocates that design into a biomedical regime through synthetic pretraining, mixed post-training, and architecture variants that trade off cross-attentive interaction against label scalability (Yazdani et al., 1 Apr 2025). Its importance lies less in any single benchmark result than in establishing a practical open-vocabulary biomedical NER template: efficient, extensible, and explicitly designed for entity schemas that need not be known in advance.