RigoBERTa-Clinical

RigoBERTa Clinical

RigoBERTa Clinical is a state-of-the-art clinical encoder language model for Spanish, developed through domain-adaptive pretraining on the largest publicly available Spanish clinical corpus, ClinText-SP. This model significantly improves performance on multiple clinical NLP benchmarks while offering robust language understanding in the clinical domain.

Model Details

Model Description

RigoBERTa Clinical was built by further pretraining the general-purpose RigoBERTa 2 on a meticulously curated clinical corpus. The pretraining leverages masked language modeling (MLM) to adapt the model’s linguistic knowledge to the Spanish clinical domain.

• Developed by: IIC
• Model type: Encoder
• Language(s) (NLP): Spanish
• License: rigoclinical-nc (permissive Non Commercial)
• Finetuned from model: RigoBERTa 2

Model Sources

• Paper: Advancing Spanish clinical language understanding through domain-adaptive pretraining and new open clinical resources

Intended Use & Limitations

Intended Use

RigoBERTa Clinical is designed for:

• Clinical text understanding in Spanish.
• Applications in healthcare NLP tasks such as clinical note classification, entity recognition in clinical texts, and related downstream tasks.
• Research and development purposes, including benchmarking and further model adaptation.

Limitations & Caveats

• Domain Specificity: Although highly effective for Spanish clinical texts, the model may not generalize to other domains or languages.
• Data Biases: ClinText-SP, while the largest corpus available, may contain biases due to source selection and the inherent limitations of public clinical data.
• Operational Cost: Despite being an encoder-based model with relatively lower computational costs compared to generative LLMs, deployment in resource-constrained settings should be carefully evaluated.

Training Details

Training Data: ClinText-SP

ClinText-SP is the largest open Spanish clinical corpus and includes data from various open sources:

• Volume: ~26 million tokens, 35,996 samples
• Sample Details: Average of ~700 tokens per sample; contains both long-form clinical cases and shorter, schematic texts,
• Sources: Medical journals, clinical shared tasks, radiological reports, and Wikipedia extracts.
• Availability: ClinText-SP on Hugging Face Datasets

Training Procedure

Preprocessing

• Tokenizer: Uses the tokenizer from RigoBERTa 2 to ensure consistency with the base model.
• Handling Long Sequences: Clinical texts exceeding 512 tokens are segmented with a stride of 128 tokens; shorter sequences are padded as necessary.
• OOV Handling: Out-of-vocabulary words are managed using subword tokenization, maintaining robust handling of clinical terminology.

Training Details

• Objective: Masked Language Modeling (MLM)
• Epochs: 2 full epochs (with the best model selected after ~1.8 epochs, based on downstream performance)
• Hyperparameters Grid:
  • Batch Sizes: 32, 64, 128
  • Learning Rates: Ranges of {5e-6, 1e-5, 2e-5} for batch size 32, {1e-5, 2e-5, 4e-5} for 64, and {1e-5, 4e-5, 8e-5} for 128
• Best Settings: Batch size = 32, Learning rate = 2e-5, ~2800 training steps (~1.8 epochs)
• Optimizer: AdamW with weight decay of 0.1
• Hardware: Trained on a single NVIDIA A100 GPU (80GB memory)

Evaluation

RigoBERTa Clinical was evaluated on several Spanish clinical NLP tasks including Named Entity Recognition (NER) and multilabel classification. Evaluation metrics (F1 score and micro-averaged F1) indicate that the model outperforms previous clinical and general Spanish language models.

Key Results:

• Achieves top performance on datasets such as cantemist, meddocan, and livingner1, among others.
• Consistently surpasses the performance of models that were trained solely on clinical data, demonstrating the advantage of leveraging general domain knowledge during domain adaptation.
• Detailed benchmarking results and comparisons are provided in the associated publication.

For a full breakdown of results (including performance on multilingual baselines and other clinical-specific models), please refer to Table 1 and the Nemenyi plot in the original paper.

Citation

If you use RigoBERTa Clinical in your research, please cite the associated paper:

BibTeX:

@article{SUBIES2026114998,
title = {Advancing Spanish clinical language understanding through domain-adaptive pretraining and new open clinical resources},
journal = {Engineering Applications of Artificial Intelligence},
volume = {178},
pages = {114998},
year = {2026},
issn = {0952-1976},
doi = {https://doi.org/10.1016/j.engappai.2026.114998},
url = {https://www.sciencedirect.com/science/article/pii/S0952197626012819},
author = {Guillem García Subies and Álvaro Barbero Jiménez and Paloma Martínez Fernández},
}

APA:

Subies, G. G., Barbero Jiménez, Á., & Martínez Fernández, P. (2026). Advancing Spanish clinical language understanding through domain-adaptive pretraining and new open clinical resources. *Engineering Applications of Artificial Intelligence, 178*, 114998. [https://doi.org/10.1016/j.engappai.2026.114998](https://doi.org/10.1016/j.engappai.2026.114998)

Model Card Authors and Contact

Guillem García Subies: [email protected], [email protected]