ReXVQA-SDFT-gemma-4-E2B

ReXVQA-SDFT-gemma-4-E2B

ReXVQA-SDFT-gemma-4-E2B is a LoRA/PEFT adapter trained on the ReXVQA dataset using Self Distillation Fine-Tuning (SDFT) for Gemma 4 E2B. It is intended for chest X-ray visual question answering and image interpretation experiments.

The adapter shows greatly improved performance on chest X-ray image interpretation, placing 2nd on the ReXrank Challenge V2.0 (rexrank.ai) with a score near 0.89. This run has not yet been submitted for official validation.

Model Details

• Adapter type: LoRA
• Base model: unsloth/gemma-4-E2B-it
• Task: Chest X-ray visual question answering / image-text-to-text generation
• Training method: Self Distillation Fine-Tuning (SDFT)
• Training data: ReXVQA
• PEFT version: 0.19.1
• License: Apache 2.0

This repository contains adapter weights and tokenizer/processor configuration. It does not contain the full base model weights.

Intended Use

This adapter is intended for research and demonstration of multimodal chest X-ray interpretation with Gemma 4 E2B. It can be loaded with the base model through transformers and peft.

It is not intended for clinical diagnosis, medical decision-making, triage, or any use where model output could directly affect patient care. Outputs should be reviewed by qualified clinicians and treated as experimental.

Evaluation

Current informal challenge result:

• Benchmark: ReXrank Challenge V2.0
• Rank: 2nd
• Score: approximately 0.89
• Validation status: not yet submitted for official validation

Additional details will be added after formal validation and reproducible evaluation artifacts are available.

Loading

import torch
from peft import PeftModel
from transformers import AutoProcessor, Gemma4ForConditionalGeneration

base_model_id = "unsloth/gemma-4-E2B-it"
adapter_id = "carlesonielfa/ReXVQA-SDFT-gemma-4-E2B"

processor = AutoProcessor.from_pretrained(adapter_id)
model = Gemma4ForConditionalGeneration.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

Limitations

• The model may produce incorrect, incomplete, or overconfident medical statements.
• Performance outside the ReXVQA/ReXrank setting has not been fully characterized.
• Dataset composition, imaging protocol differences, and prompt wording may affect output quality.
• This adapter inherits limitations and safety considerations from the base Gemma 4 E2B model.

Citation

If you use this adapter, please cite the relevant ReXVQA/ReXrank resources and the Gemma base model according to their respective citation guidance.