yasserrmd/OphthaScholar-1.2B

🧠 OphthaScholar‑1.2B

OphthaScholar‑1.2B is a fine-tuned clinical assistant model focused specifically on ophthalmology. It is built on top of the LiquidAI LFM‑1.2B base model and fine-tuned using ophthalmology-relevant question-answer pairs extracted from the MIRIAD‑4.4M dataset.

---

📌 Model Use

• ✔️ Good at: Direct factual question answering related to ophthalmology (short-form)
• ⚠️ Not suitable for: Multi-choice QA or reasoning-based multi-step inference
• ❌ Do not use for: Real clinical decision-making or diagnostic support

---

⚠️ Important Caution (from MIRIAD Dataset)

This model was trained using a filtered subset of the MIRIAD‑4.4M dataset. While the dataset was curated using LLMs and rule-based methods, no licensed medical professionals reviewed it. As noted by the dataset authors:

“Despite quality filtering, hallucinations or factual errors may remain. The dataset should not be used to train clinical decision tools or models intended for real medical use.”

---

🔧 How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "yasserrmd/OphthaScholar"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, 
             device_map="auto", 
             torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
             trust_remote_code=True).eval()

question = "What is the treatment for neovascular age-related macular degeneration?"
prompt = f"System: You are an ophthalmology expert.\nUser: {question}\nAnswer:"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=64, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True).split("Answer:")[-1].strip())

---

🧠 Suggested Applications

• ✅ Clinical education (ophthalmology)
• ✅ Dataset generation for supervised fine-tuning
• ✅ Question answering for medical research chatbots

---

❌ Not Intended For

• ❌ Clinical use
• ❌ Diagnostic automation
• ❌ Emergency decision tools

---

📖 Citation

@misc{ophthascholar2025,
  title={OphthaScholar-1.2B: A factual ophthalmology model fine-tuned on MIRIAD},
  author={Yasserrmd},
  year={2025},
  url={https://huggingface.co/yasserrmd/OphthaScholar}
}

---

📎 License

• Base: MIT (Liquid LFM-1.2B)
• Data: MIRIAD-4.4M (filtered use only; not for clinical deployment)

---