KnowMedPhi3.5-mini

KnowMedPhi3.5-mini is a medical language model developed from Phi-3.5-mini using expert knowledge-centred continued pre-training followed by medical instruction tuning.

The model accompanies the study:

KnowMedPhi: Expert knowledge-centred pre-training for transferable clinical language modelling

Model: https://huggingface.co/knowlab-research/KnowMedPhi3.5-mini
Project repository: https://github.com/knowlab/KnowMedPhi
Paper: [PAPER LINK]

Model description

Medical language models are commonly adapted through continued pre-training on large collections of biomedical literature. KnowMedPhi investigates an alternative strategy in which continued pre-training is performed on a substantially smaller corpus centred on explicitly encoded biomedical and clinical knowledge.

KnowMedPhi3.5-mini was developed in two stages:

1. Continued pre-training: Phi-3.5-mini was further pre-trained using KnowMed-PT, an approximately 126-million-token corpus constructed from biomedical and clinical knowledge resources.
2. Medical instruction tuning: The resulting model was subsequently instruction-tuned using KnowMed-IT, a corpus of 365,547 medical instruction examples.

The KnowMed-PT and KnowMed-IT resources are documented separately in the project repository.

Base model

• Base architecture: Phi-3.5-mini
• Continued pre-training corpus: KnowMed-PT
• Instruction-tuning corpus: KnowMed-IT
• Training framework: LLaMAFactory
• Distributed training: DeepSpeed

Further training details are reported in the accompanying manuscript.

---

Training data

KnowMed-PT

KnowMed-PT is the expert knowledge-centred corpus used for continued pre-training. It contains approximately 126 million tokens assembled from biomedical and clinical knowledge resources.

Source	Approx. tokens
AGCT / SNOMED CT-derived data	33.4M
PMC-Patients	8.1M
Medical Wikipedia data	15.9M
DrugBank	4.5M
PathBank	62.3M
MONDO	1.2M
Human Phenotype Ontology (HPO)	0.8M
Total	126.2M

KnowMed-PT: Dataset

The constituent resources remain subject to their respective licences and terms of use. Where redistribution of source-derived content is restricted, users should obtain the corresponding resource from its original provider and comply with the applicable licence.

Detailed source provenance and access information are provided in the accompanying dataset documentation.

---

KnowMed-IT

KnowMed-IT is the medical instruction-tuning corpus used following continued pre-training. It contains 365,547 instruction examples assembled from multiple medical instruction, question-answering, and educational resources.

Source	Examples
Asclepius	158,114
AlpaCare / MedInstruct	52,002
Medical-QA	41,992
MedMCQA rationales	36,316
NHS QA / OpenGPT	29,354
MedlinePlus-derived instructions	20,891
MedQuAD	16,407
MedQA rationales	9,816
NEJM-AI-Exams	655
Total	365,547

KnowMed-IT: Dataset

The MedlinePlus-derived instruction data were constructed as part of this work. Other constituent datasets remain subject to the licences and terms of their original sources.

Where required, only the designated training splits of benchmark-derived datasets were used for instruction tuning.

Evaluation

KnowMedPhi3.5-mini was evaluated across biomedical and clinical language tasks covering:

• medical multiple-choice question answering;
• biomedical named entity recognition;
• biomedical document classification;
• clinical natural language inference;
• biomedical question answering;
• biomedical multi-hop reasoning;
• medical explanation;
• fact verification;
• treatment planning; and
• information extraction.

The public benchmark datasets and their original sources are listed in the project repository.

The accompanying study additionally evaluates transfer to structured information extraction from de-identified histopathology reports from five NHS Trusts. These clinical data are not publicly distributed because they are governed by NHS information-governance, data-sharing, and institutional access requirements.

Intended use

KnowMedPhi3.5-mini is intended for research on biomedical and clinical natural language processing, medical language-model evaluation, and continued-pre-training strategies.

The model is not intended to replace qualified healthcare professionals or to provide autonomous diagnosis, treatment recommendations, or other clinical decisions.

Limitations

KnowMedPhi3.5-mini may generate incorrect, incomplete, unsupported, or outdated medical information. Performance varies across tasks and clinical domains, and benchmark performance should not be interpreted as evidence of safety or effectiveness in clinical deployment.

Clinical use requires appropriate task-specific validation, governance, human oversight, and assessment within the intended deployment setting.

Citation

If you use KnowMedPhi3.5-mini in your work, please cite:

@article{kim2026knowmedphi,
  title   = {KnowMedPhi: Expert knowledge-centred pre-training for transferable clinical language modelling},
  author  = {Kim, Yunsoo and others},
  year    = {2026}
}