VaidenceVLM-1.0 is a research-only PEFT/QLoRA adapter for OpenGVLab/InternVL3_5-4B. This repository contains adapter weights, not a standalone 4B base model. It requires the named base model, a CUDA GPU, and trust_remote_code=True.

It is not a medical device and must not be used for diagnosis, triage, treatment, or patient care.

Quick start

bash

git clone https://huggingface.co/dipankar500/VaidenceVLM-1.0
cd VaidenceVLM-1.0
python -m pip install -r requirements.txt
python inference.py --adapter . --image /path/to/image.jpg --question "Describe visible findings and state uncertainty."

SYSTEM_PROMPT.txt is a safety-oriented interactive prompt. It was not used for the reported benchmark scores, so do not prepend it when reproducing those results.

Reproducible generative evaluation — not official leaderboard metrics

Benchmark / split	Questions	Normalized exact match	Token F1	Yes/No exact match	Errors
SLAKE English test	1061	57.49%	62.33%	80.85%	0
VQA-RAD test	32	53.12%	60.64%	77.78%	0

Metrics use our released greedy generative evaluator and custom normalized answer matching. They are not official benchmark leaderboard scores, clinical accuracy, evidence of state of the art, or a comparison with Gemini or another proprietary model. See EVALUATION.md.

Repository contents

adapter_model.safetensors, adapter_config.json: PEFT adapter weights/configuration.
tokenizer/: recovered tokenizer files.
inference.py: verified local inference entry point.
SYSTEM_PROMPT.txt: default interactive research-safety prompt.
EVALUATION.md, TRAINING_SUMMARY.md, LIMITATIONS.md: transparent release documentation.

Citation

See CITATION.cff. Cite InternVL 3.5 and the original datasets in research use.