abhijit2k01/medgaze-adapters

MedGaze adapters

Decode-only, rank-8 LoRA adapters on the query rows of each model's top-100 fixation heads (the attention heads that look where a question points), trained on SLAKE closed questions with an answer loss plus a mask-supervised attention term (recipe "V-g", 400 steps).

import os
from huggingface_hub import snapshot_download
os.environ["GAZEMED_ADAPTERS"] = snapshot_download("abhijit2k01/medgaze-adapters")
from medgaze import GazeModel
gm = GazeModel.from_pretrained("Qwen/Qwen3-VL-2B-Instruct", gaze="auto")
print(gm.describe())

index.json lists every adapter with its checksum, what it was measured to do and its caveats (the accuracy gain is in-domain; it does not transfer to VQA-RAD). rankings/ holds the fixation-head rankings the adapters were trained on. Research use only; not a medical device.

Paper, code and project page: https://aj-das-research.github.io/GazeHeads-Medical-VLMs/