IndigoAtlas25/lotus-mini

Lotus Mini

A sentiment-routed medical Mixture-of-Experts: a tiny sentiment classifier routes each patient query to one of two LoRA expert adapters, all sharing a single 4-bit quantized Gemma 3 4B base (QLoRA).

user prompt
  |--> DistilBERT sentiment classifier (ONNX)
  |      P(POS) >= 0.5 -> POS expert   (grateful/reassured patients)
  |      else          -> NEG expert   (worried/anxious patients)
  v
shared 4-bit Gemma 3 4B + chosen LoRA expert
  v
medical answer

Contents of this repo

Path	What
experts/neg	LoRA adapter trained on negative-sentiment patient queries
experts/pos	LoRA adapter trained on positive-sentiment patient queries
router/	DistilBERT SST-2 sentiment classifier (ONNX) used as the router

How to use (MLX / Apple Silicon)

The adapters are trained on top of mlx-community/gemma-3-4b-it-4bit. Load the base model and attach the routed adapter:

from mlx_lm import load
from mlx_lm.utils import load_adapters

model, tokenizer = load("mlx-community/gemma-3-4b-it-4bit")
load_adapters(model, "lotus-mini/experts/neg")   # or .../experts/pos

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "I've had chest pain for three days and I'm scared."}],
    add_generation_prompt=True,
)

Or clone the source repo (has the full pipeline, setup script, and trainer):

git clone https://github.com/2028badivi/lotus-mini.git && cd lotus-mini
./scripts/setup.sh
uv run moe.py "I've had chest pain for three days and I'm scared."

Training

• Base: mlx-community/gemma-3-4b-it-4bit (4-bit QLoRA base, ~3.4 GB)
• Data: lavita/medical-qa-datasets (config chatdoctor_healthcaremagic, 112k doctor-patient Q&A pairs), stratified by the sentiment of the patient's description into POS/NEG buckets
• Adapters: rank-8 LoRA on 8 of 34 layers, 80 iterations each, ~14 MB each

Disclaimer

This is a research/demo model fine-tuned from AI-generated medical Q&A data. It is not a substitute for professional medical advice, diagnosis, or treatment.