Examples

GAZE includes five complete example applications demonstrating different use cases.

NOVA brain MRI

Location: examples/nova/ (README)

Brain MRI analysis with three sub-tasks: caption generation, diagnosis prediction, and lesion localization. Uses the NOVA dataset which auto-downloads from HuggingFace.

pip install gaze-vlm[nova]

# Single-turn mode
uv run python -m examples.nova.src.cli \
  --model openai/gpt-4o \
  --mode single_turn \
  --max-samples 10

# Agentic mode with tools
uv run python -m examples.nova.src.cli \
  --model openai/gpt-4o \
  --mode agentic \
  --use-tools \
  --max-turns 5 \
  --max-samples 10

GEMeX visual grounding

Location: examples/gemex_thinkvg/ (README)

Visual grounding with chain-of-thought reasoning on chest X-rays. Requires MIMIC-CXR access (PhysioNet credentialed).

pip install gaze-vlm[gemex]

uv run python -m examples.gemex_thinkvg.eval \
  --dataset ./data/test.jsonl \
  --image-dir /path/to/mimic-cxr-jpg \
  --model openai/gpt-4o \
  --mode agentic \
  --use-tools \
  --output ./results

AgentClinic NEJM

Location: examples/agentclinic_nejm/ (README)

Multi-turn diagnostic reasoning where the model gathers clinical information (history, exam, tests, imaging) before making a diagnosis.

pip install gaze-vlm[agentclinic]

uv run python -m examples.agentclinic_nejm.eval \
  --dataset ./data/agentclinic_nejm_extended.jsonl \
  --model openai/gpt-4o \
  --num-samples 10 \
  --output ./results

PubMedQA

Location: examples/pubmedqa/ (README)

Text-only medical Q&A with yes/no/maybe answers. Uses the PubMedQA dataset (auto-downloads).

pip install gaze-vlm[pubmedqa]

uv run python -m examples.pubmedqa.src.cli \
  --model openai/gpt-4o \
  --mode single_turn \
  --max-samples 50

VQA-RAD

Location: examples/vqa_rad/ (README)

Radiology visual question answering with closed and open-ended questions. Uses the VQA-RAD dataset (auto-downloads).

pip install gaze-vlm[vqa-rad]

uv run python -m examples.vqa_rad.src.cli \
  --model openai/gpt-4o \
  --mode agentic \
  --use-tools \
  --max-samples 20

Local models

All examples support local model inference via LM Studio. Pass --base-url to point at your instance:

uv run python -m examples.pubmedqa.src.cli \
  --model qwen3.5-a3b \
  --base-url http://localhost:1234/v1 \
  --mode single_turn \
  --max-samples 5

Writing your own example

See Getting started for the pattern. The examples vary in size, but the NOVA example (examples/nova/src/) is representative. Note that evaluation/ is a package, not a single module, and dataset loading lives under data/:

examples/your_task/
    src/
        __init__.py
        cli.py           # CLI entry point (argparse + run_evaluation)
        config.py        # Frozen config dataclass for the task
        processor.py     # AgenticProcessorBase subclass
        schemas.py       # Response schema(s) and validate_response()
        data/            # Dataset loading package
            __init__.py
        evaluation/      # Metrics package, one module per sub-task
            __init__.py
            caption.py
            detection.py
            diagnosis.py
    run_local.sh         # LM Studio convenience script
    README.md

Smaller examples (for instance pubmedqa) collapse some of these into fewer modules. The only hard requirement is a processor that subclasses AgenticProcessorBase and implements the four abstract methods.