medichat-flan-t5-base-lora

MediChat – FLAN-T5 Base LoRA Adapter

A lightweight LoRA adapter trained on top of google/flan-t5-base to add medical-assistant style capabilities. This repository contains only the adapter weights and tokenizer files needed to reproduce the exact behaviour during inference.

📁 Repository Contents

This model repo includes the following files:

File Purpose adapter_config.json Configuration for the LoRA adapter (rank, alpha, target modules, etc.). adapter_model.safetensors LoRA adapter weights in safetensors format. tokenizer.json Full tokenizer definition. tokenizer_config.json Tokenizer settings and metadata. special_tokens_map.json Mapping for special tokens (PAD, EOS, etc.). spiece.model SentencePiece model used by FLAN-T5. .gitattributes Repo attributes. README.md Documentation.

Note: This adapter must be loaded on top of the base model google/flan-t5-base.

🔧 Intended Use

The adapter specializes FLAN-T5 for:

Medical question answering

Symptom explanation

Short clinical-style responses

Health-related conversational tasks

It is not a medical authority. All outputs should be double-checked by qualified professionals.

⚠️ Limitations & Safety

May generate incorrect or incomplete medical information.

Should not be used for diagnosis or emergency guidance.

Training data details are not provided here, so biases may exist.

The base model itself is not medically specialized.

Use responsibly and always verify outputs.

🚀 Quick Start (Python)

Install requirements:

pip install transformers accelerate peft safetensors

Load the base model + LoRA adapter:

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM from peft import PeftModel

BASE = "google/flan-t5-base" ADAPTER = "wilsonl/medichat-flan-t5-base-lora"

Load tokenizer

tokenizer = AutoTokenizer.from_pretrained(BASE)

Load base model

model = AutoModelForSeq2SeqLM.from_pretrained(BASE, device_map="auto")

Apply the LoRA adapter

model = PeftModel.from_pretrained(model, ADAPTER)

Run inference

prompt = "A patient asks: I have a sore throat and fever. What should I do?" inputs = tokenizer(prompt, return_tensors="pt").to(model.device) output = model.generate(**inputs, max_new_tokens=200)

print(tokenizer.decode(output[0], skip_special_tokens=True))

🧪 Example Prompts

Short clinical advice

Patient: "I feel dizzy when standing up quickly." Give a short explanation and two safety tips.

Case summarization

Summarize the following case in 3 bullet points: 36-year-old with fever, mild cough, no allergies, using paracetamol.

Medical assistant style

You are a helpful medical assistant. Respond concisely and clearly. User: What causes chest tightness after exercise?

🏗️ How This Adapter Was Created

This repository contains only the LoRA adapter weights. Training details such as:

dataset

preprocessing

hyperparameters

epochs

evaluation metrics

are not included in this model card and should be added later if needed.

📜 License

No explicit license is included in this repository. Check with the model owner before using the model commercially.

🤝 Contributing

You can extend this repo by adding:

A full training summary

Evaluation metrics

A changelog

An inference notebook

A detailed model card

📬 Contact

For questions or modifications, reach out via the HuggingFace model page: wilsonl/medichat-flan-t5-base-lora