Jibay2-Medical-v2

A fine-tuned version of JibayAi/Jibay_2 that converts messy, shorthand clinical notes into clean, structured patient summaries.

Model Details

---

Intended Use

Primary Use

This model helps healthcare professionals quickly understand a patient's background by transforming fragmented clinical shorthand (e.g., pt 58 yo M c/o chest pain x2d hx HTN DM2) into a clean, structured summary organized into standardized sections:

• Patient Background
• Chief Complaint
• Medical History
• Medications
• Vital Signs
• Examination
• Investigations
• Assessment

Intended Users

• Clinical researchers
• Medical documentation teams
• Healthcare AI developers
• Medical education platforms

Out-of-Scope Uses

• Not a diagnostic tool. The model does not diagnose, treat, or prescribe.
• Not for direct clinical decision-making. All outputs must be reviewed by a qualified healthcare professional.
• Not for processing real patient data without de-identification. Using PHI without proper safeguards may violate HIPAA/GDPR.
• Not multilingual. Trained only on English clinical text.

---

Training Data

Data Sources

• TheBlueScrubs-v1-fixed (openmed-community) — filtered medical text corpus
• Hindawi medical journals (mkurman/hindawi-journals-2007-2023) — open-access medical literature

Data Pipeline

1. Raw medical text was filtered to identify genuine clinical notes using a keyword-based heuristic (~2,645 from 70,000 total examples).
2. Each clinical note was converted into messy shorthand using a rule-based transformation (medical abbreviations, structure removal, connective tissue stripping).
3. Qwen2.5-7B-Instruct generated clean, structured summaries from the messy notes, producing a parallel corpus of {messy, clean} pairs.
4. Pairs passing quality filters were retained.

Dataset Statistics

• Average messy input length: ~900 characters
• Average clean output length: ~803 characters

Data Limitations

• Rule-based "messy" transformation may not perfectly reflect real-world clinical shorthand.
• Source corpus contains heterogeneous medical content (case reports, journal abstracts, patient notes).
• No formal medical validation of generated summaries was performed.

---

Training Procedure

Training Hyperparameters

Training Framework

• Unsloth for 2x faster fine-tuning and reduced memory usage
• TRL SFTTrainer for supervised fine-tuning
• PEFT for LoRA adapters
• BitsAndBytes for 4-bit quantization

Hardware

• Google Colab Pro
• NVIDIA A100 (40 GB)

Evaluation Strategy

• Validation loss evaluated every 20 steps
• Early stopping with patience of 3 evaluations
• Best checkpoint loaded based on lowest validation loss

Training Results

• Best validation loss: 1.4403
• Training time: ~5.7 minutes

---

Usage

Loading the Model

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

BASE_MODEL = "JibayAi/Jibay_2"
ADAPTER = "aradmanamnaoon/jibay2-medical-v2"

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
base = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    torch_dtype=torch.float16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, ADAPTER)
model = model.merge_and_unload()

Running Inference

prompt = '### Input:\n45yo F abd pain 2d nausea no vomit.\n\n### Clean Summary:\n'

inputs = tokenizer(prompt, return_tensors='pt').to(model.device)
outputs = model.generate(**inputs, max_new_tokens=600, temperature=0.7)
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result.split('### Clean Summary:')[-1].strip())

Expected Output Format

## Patient Background
- 45-year-old female

## Chief Complaint
- Abdominal pain for 2 days, associated with nausea without vomiting

## Medical History
- Appendectomy (2019)

## Vital Signs
- Temperature: 37.8C
- Blood pressure: 120/80 mmHg

## Examination
- Abdomen: Soft, tender in the right lower quadrant

---

Limitations

Known Limitations

• Hallucination risk. The model may invent clinical details not present in the input. Always verify outputs against the source note.
• Not clinically validated. No medical professional has reviewed the model's outputs for accuracy.
• Training data bias. The model inherits biases from the underlying corpus.
• Limited to English. Does not support multilingual clinical notes.
• Rule-based messy transformation. Messy inputs during training were generated procedurally, not from real clinical shorthand.
• Small training set. 2,634 examples is small; performance may degrade on out-of-distribution inputs.

Ethical Considerations

• Human-in-the-loop required. This model is a documentation aid, not a replacement for clinical judgment.
• PHI handling. Do not feed protected health information without explicit consent and appropriate safeguards.
• No clinical deployment. Not cleared or approved by any regulatory body (FDA, EMA, MHRA, etc.).

Bias and Fairness

• Not evaluated for demographic fairness across age, sex, ethnicity, or socioeconomic groups.
• Performance may vary significantly across patient populations.

Environmental Impact

Training used a single NVIDIA A100 GPU for approximately 6 minutes.

Citation

@misc{jibay2-medical-v2,
  title = {Jibay2-Medical-v2: Fine-tuned Medical Note Summarizer},
  author = {aradmanamnaoon},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/aradmanamnaoon/jibay2-medical-v2}}
}

Acknowledgments

• JibayAi for the base Jibay 2 model
• Unsloth for the training framework
• Qwen team for Qwen2.5-7B-Instruct used in data generation
• Hugging Face for model hosting

---

Disclaimer: This model is for research purposes only. It is not a medical device and must not be used for clinical decision-making without review by a qualified healthcare professional.