🧠 Relational AI (v1)
This is a specialised Large Language Model (LLM) fine-tuned to understand and generate person-centred nursing documentation. It has been trained on the principles of relational care as described in the academic archives of the Foundation of Nursing Studies (FoNS) and the International Practice Development Journal (IPDJ).
The model is designed to assist nurses, students, and clinical educators in moving from messy, real-world transcripts to structured, evidence-based, and compassionate clinical notes.
Model Details
- Developed by: Lincoln Gombedza (Nurse Citizen Developer, Registered Learning Disability Nurse & Practice Educator
- Model Type: Causal Language Model
- Language: English (Clinical Nursing)
- License: MIT
- Finetuned from model:
microsoft/phi-3-mini-4k-instruct
🎆 Clinical Leadership & Vision
This model is a technical artifact designed to support the work of clinical leaders in Health Equity and Relational Care:
- Health Equity: The model's development aligns with the principles championed by Kumbi Kariwo (Nurse Citizen Developer & Health Inequalities Lead).
- Relational Care: The model's focus on person-centredness is inspired by the vision of Professor Joanne Bosanquet and Dr. Clare Cable.
How to Get Started with the Model
Use this model with the transformers library to summarise and structure clinical transcripts.
from transformers import AutoModelForCausalLM, AutoTokeniser, pipeline
# Load the model you created
model_name = "NurseCitizenDeveloper/fons-relational-ai-v1"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
# Create a text generation pipeline
pipe = pipeline("text-generation", model=model, tokeniser=tokeniser)
# Example: A messy transcript
transcript = """
Hi Dave, checking on you. Leg hurts? An 8? Ouch. Alright, pulse is 98, BP is 140 over 90.
I'm going to get you some Oramorph. Back in a sec... okay it's taken. I'll be back in 30.
"""
# The prompt format must match the training data
prompt = f"""### Instruction:
Summarize this excerpt from a nursing journal and explain its core principle for person-centred nursing practice.
### Input:
{transcript}
### Response:
"""
# Generate the structured note
result = pipe(prompt, max_new_tokens=200, num_return_sequences=1)
print(result[0]['generated_text'])
Expected Output: The model will produce a structured SBAR note, similar to the examples used during training.
Training Details
Training Data
The model was fine-tuned on a synthetic dataset of ~18,000 examples. These examples were generated by processing the text of 378 peer-reviewed articles from the Foundation of Nursing Studies (FoNS) / IPDJ library. Each example consists of a text chunk from an article and an instruction to summarize its core principle, forcing the model to learn the language and concepts of relational care.
Training Procedure
The model was fine-tuned using Supervised Fine-Tuning (SFT) with the trl library. We used the Unsloth framework for 4-bit quantization (QLoRA) to enable training on a free Google Colab T4 GPU.
max_steps: 150 per_device_train_batch_size: 2 gradient_accumulation_steps: 4 learning_rate: 2e-4
Uses and Limitations
Direct Use
This model is intended as a documentation assistant and educational tool. It can be used to:
- Draft reflective accounts for NMC Revalidation.
- Convert messy transcripts into structured SBAR notes.
- Help student nurses learn the language of person-centred practice.
Out-of-Scope Use
This model is NOT a medical device. It is not intended for direct clinical decision-making, diagnosis, or treatment. All outputs must be reviewed and validated by a registered clinician.
Bias, Risks, and Limitations
This model is an experimental, first-generation artifact (v1). Users (both direct and downstream) should be aware of the following risks and limitations.
⚠️ CRITICAL FINDING (v1): Repetitive Degeneration
Issue: The current version of the model (v1) has a tendency to degenerate into repetitive, nonsensical text loops (e.g., "tu tu tu tu...", "becausebecausebec...") after generating approximately 150-200 tokens.
Root Cause Analysis:
This is an issue with the fine-tuning process, not the prompt. The training dataset, while embodying correct clinical principles, may contain formatting artifacts or a lack of diversity that leads to this degenerative behavior in the base model (phi-3-mini). Standard mitigation parameters (repetition_penalty, temperature) are insufficient to fully correct this.
Recommended Solutions:
- Immediate Workaround: When using this model in a pipeline, set
max_new_tokensto a value between 150 and 200 to truncate the output before the repetition begins. - Model-Level Fix (Planned for v2):
- Data Curation: The fine-tuning dataset will be reviewed for repetitive patterns and corrupted entries.
- Data Augmentation: We will increase the diversity and volume of the training data.
- Hyperparameter Tuning: We will experiment with different learning rates and training epochs.
Other Limitations
- Knowledge Cutoff: The model's knowledge is limited to the content of the FoNS articles it was trained on. It does not have real-time clinical knowledge.
- Hallucination: While fine-tuning reduces the risk, the model may still generate plausible but incorrect information.
- Inherited Bias: The model inherits the limitations and potential biases of its base model and the training data.
Paper: From Analogue Wisdom to Digital Practice - Codifying Relational Care
Abstract
The digitisation of healthcare risks creating "transactional" records that capture clinical tasks but erase the relational work central to nursing. This paper details the creation of fons-relational-ai-v1, a specialized Large Language Model (LLM) designed to bridge this gap. By fine-tuning the microsoft/phi-3-mini architecture on a synthetic dataset derived from the principles of the Foundation of Nursing Studies (FoNS), we have developed an open-source tool that can understand, validate, and generate person-centred nursing documentation. This work represents a critical step in ensuring that as AI enters clinical practice, it is grounded in the evidence-based wisdom of nursing theory and actively promotes health equity.
1. Introduction: The Analogue-to-Digital Gap
Nursing practice is a blend of clinical science and relational art. However, as we move from analogue (pen and paper) to digital (EHRs and AI scribes), the "art" is often lost in translation. Standard AI models, trained on generic medical data, are proficient at structuring facts but fail to capture the nuance of empathy, active listening, or shared decision-making. This creates a critical unmet need for AI that speaks the language of nursing.
This project, led by a Nurse Citizen Developer, aimed to prove that nursing's core values can be codified into AI from the ground up, not as an afterthought.
2. Methodology: Building the "Wisdom Engine"
The creation of this model followed a four-stage process:
a) Evidence Harvesting: We developed a Python-based web scraper to download 378 peer-reviewed articles from the open-access archives of the International Practice Development Journal (IPDJ), provided by the Foundation of Nursing Studies. This formed our evidence base.
b) Synthetic Data Generation: Due to the Non-Commercial license of the source material, we did not use the text directly for training. Instead, we used Azure OpenAI (GPT-4o) to generate a synthetic dataset of ~18,000 "flashcards." Each flashcard contained a messy, realistic clinical transcript and a "perfect" SBAR note that explicitly applied a core principle (e.g., "Relational Care," "Health Equity") found within the FoNS library.
c) Supervised Fine-Tuning (SFT): Using a Google Colab GPU, we fine-tuned the microsoft/phi-3-mini-4k-instruct model on this synthetic dataset. The process utilized the Unsloth library for 4-bit QLoRA, making it possible to train a high-quality model with limited resources.
d) Publication: The resulting model was published to the Hugging Face Hub under an MIT license to ensure maximum accessibility for the global nursing and developer communities.
3. Health Equity by Design (The Coded Bias Protocol)
A core objective was to build an actively anti-racist AI. Standard models perpetuate bias by relying on outdated metrics like the Fitzpatrick scale. Inspired by the leadership of Kumbi Kariwo (Health Inequalities Lead, BCHC), our training data explicitly teaches the model to:
- Reject "redness" as a sole indicator of pressure damage on darker skin tones (Monk Scale MST-06+).
- Demand tactile assessment descriptors (heat, bogginess, induration) for equitable wound care.
4. Conclusion: Empowering the Nurse Citizen Developer
fons-relational-ai-v1 is more than a model; it is a proof of concept. It demonstrates that nurses, armed with modern tools and deep clinical expertise, can and must be the architects of the AI that will shape our future practice. By making this tool open source, we invite a global community of Nurse Citizen Developers to contribute, critique, and build upon this foundation, ensuring the digital future of nursing is safe, equitable, and profoundly human.
References
- Foundation of Nursing Studies (FoNS) & IPDJ: https://www.fons.org/library/journal
- Clinical Leadership: Kumbi Kariwo (BCHC), Dr. Clare Cable (QNIS), Professor Joanne Bosanquet (FNS).
Citation
If you use this model in your research, please cite it using its Digital Object Identifier (DOI):