gpt-oss Endocrinology Scenario-Based MCQ Models (8 Variants)
This collection contains 8 instruction-tuned gpt-oss variants (20B & 120B) optimized for endocrinology scenario-based MCQs and clinical Q&A. Variants are provided as LoRA adapters, merged FP16, and GGUF builds (FP16 and Q4_K_M) for flexible deployment from research servers to local/offline inference.
What the models can do
All variants are designed for three primary inference modes:
- Scenario MCQ Answering (with explanation): Provide the single best answer (A/B/C/D) and explain the reasoning.
- Scenario MCQ Generation: Generate a full clinical MCQ (stem + options + correct answer + brief explanation).
- General Endocrinology Q&A: Provide concise, structured answers to general endocrinology questions.
Important: These models are intended for educational and research purposes only and must not be used for clinical decision-making.
Dataset provenance and medical review
The dataset was designed and prepared by Professor _Saeed Awad M. Alqahtani _ (Taibah University, Saudi Arabia) and reviewed and approved by three licensed medical doctors:
- Two consultant endocrinologists
- One general medical doctor
Model variants in this collection:
20B (4 variants)
- HAJERI/gpt-oss-20b-endocrinology-scenario-mcq-it-lora
- HAJERI/gpt-oss-20b-endocrinology-scenario-mcq-it-merged-fp16
- HAJERI/gpt-oss-20b-endocrinology-scenario-mcq-it-fp16-gguf
- HAJERI/gpt-oss-20b-endocrinology-scenario-mcq-it-Q4_K_M-gguf
120B (4 variants)
- HAJERI/gpt-oss-120b-endocrinology-scenario-mcq-it-lora
- HAJERI/gpt-oss-120b-endocrinology-scenario-mcq-it-merged-fp16
- HAJERI/gpt-oss-120b-endocrinology-scenario-mcq-it-fp16-gguf
- HAJERI/gpt-oss-120b-endocrinology-scenario-mcq-it-Q8_0-gguf
How to choose the right variant
Pick by deployment format
-
-lora: best if you want to keep the base model separate and apply adapters (research workflows, experimentation).
-
-merged-fp16: best for GPU inference with a standard Transformers pipeline (higher quality, higher VRAM).
-
-fp16-gguf: best for llama.cpp inference with high quality (larger RAM/VRAM than quantized).
-
-Q4_K_M-gguf: best for fast, memory-efficient local inference (most users on laptops/desktops).
Pick by model size
-
20B: The efficient specialist. Offers a strong balance of speed and medical reasoning. Capable of running on high-end consumer hardware (e.g., single RTX 3090/4090 with quantization). Ideal for rapid iteration on standard endocrinology MCQs and instruction-following tasks.
-
120B: The heavy lifter. Provides maximum reasoning depth and nuance for complex clinical vignettes. Best for ambiguous scenarios where "textbook" answers aren't enough. Requires significant compute resources (multi-GPU or enterprise-grade servers) but delivers the highest potential accuracy.
Pick by format (Bonus)
-
LoRA: Best for developers wanting to apply this fine-tune to a base model they already have cached, or for further training.
-
Merged fp16: The full-precision, standalone model. Best for future merging or when exact precision is critical (no quantization loss).
-
GGUF (Q4/Q8): Best for local deployment (llama.cpp, LM Studio). Use Q8 for near-lossless quality on high RAM, or Q4 to fit the 20B model into widely available VRAM.
Prompt templates (the 3 modes)
Mode 1 — MCQ Answer + Explanation
You are an expert endocrinologist.
Clinical scenario:
A 32-year-old woman presents with weight loss, heat intolerance, and palpitations.
Labs: suppressed TSH and elevated free T4.
Options:
A. Subacute thyroiditis
B. Graves disease
C. Hashimoto thyroiditis
D. Pituitary adenoma
Instructions:
- Choose the single best answer.
- Explain the reasoning.
Mode 2 — Generate a full MCQ
You are an expert endocrinologist.
Create a clinical multiple-choice question on diabetic ketoacidosis suitable for
final-year medical students.
Include:
- A clinical scenario
- Four answer options
- The correct answer
- A brief explanation
Mode 3 — General Endocrinology Q&A
You are an expert endocrinologist.
Question:
Explain the mechanism of action of metformin.
Inference examples
A) GGUF inference with llama.cpp (recommended for GGUF models)
Example (Q4_K_M GGUF):
./main
-m gemma-3-12b-endocrinology-scenario-mcq-it-Q4_K_M.gguf
-p "You are an expert endocrinologist. Explain the pathophysiology of diabetic ketoacidosis."
-n 400
--temp 0.7
Example (FP16 GGUF):
./main
-m gemma-3-12b-endocrinology-scenario-mcq-it-fp16.gguf
-p "You are an expert endocrinologist. Create a clinical MCQ on DKA with 4 options and explanation."
-n 500
--temp 0.7
B) Merged FP16 inference with Transformers (for -merged-fp16)
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "HAJERI/gemma-3-12b-endocrinology-scenario-mcq-it-merged-fp16"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype="auto")
prompt = """You are an expert endocrinologist.
Question:
Explain the mechanism of action of metformin.
"""
inputs = tok(prompt, return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=350, temperature=0.7, do_sample=True)
print(tok.decode(out[0], skip_special_tokens=True))
Safety note
These models can generate clinically plausible text. Always apply:
-
Human expert review
-
Educational context only
-
No patient-specific decisions
Compute acknowledgement
Compute resources (NVIDIA A100 80GB) were supplied by the National eLearning Center (NeLC), Saudi Arabia, as part of the AI Sandbox in Digital Learning initiative.
Authors and contact
Model author: Professor Saeed Awad M. Alqahtani
Affiliation: Taibah University, Saudi Arabia
For academic, educational, or research inquiries: please open an issue on the relevant model repository.
