NV-Reason-CXR Chest X-ray Reasoning Test
Runs reproducible command-shape and live inference smoke tests for chest X-ray reasoning.
The documentation clearly limits the skill to engineering validation and prohibits diagnosis and treatment use. It discloses cache writes, output-directory writes, dependency installation, and network access to Hugging Face/GitHub/hf.space, and pins the model revision. There is no evidence of malware, credential theft, or destructive defaults. Points are deducted because there is no explicit pre-execution confirmation, patient-sensitive-image handling or retention policy, consent step before remote upload, or rollback plan; local setup may also modify the active Python environment.
The entrypoint, arguments, error handling, mock path, input validation, and tests are largely consistent. Tests cover generated fixtures, missing inputs, invalid images, API option conflicts, and streaming-response parsing. Static review cannot confirm actual execution, while the live path depends on GPU availability, a remote Space, exact library versions, and an upstream Gradio schema; abnormal-path coverage remains limited, so the score is conservatively bounded by static calibration.
The audience, research/engineering scenarios, input formats, JSON output, and non-fit boundaries are clearly described, with routing rules for different request types. Points are deducted because Chinese-language interaction is not documented, the API path depends on an overseas Hugging Face service, and trigger boundaries lack broader negative examples.
SKILL.md provides purpose, prerequisites, limitations, troubleshooting, exact commands, environment variables, dependencies, versions, and upstream references. The manifest also records licensing, runtime, and verification metadata. Points are deducted because the benchmark explicitly reports a missing recommended Examples section, there is no formal changelog, and users must reconcile Apache-2.0 code terms with the noncommercial license for model weights.
The skill can directly perform command-shape smoke tests and emit JSON containing input metadata, runtime information, response text, and limitations; the mock fixture has direct value for wiring verification. Points are deducted because mock output is not evidence of model quality, live outputs may hallucinate or miss findings, clinical value is explicitly excluded, and only two internal evaluation tasks are reported without broad representative evidence or comparison with alternatives.
The files provide pinned upstream commits and model revision, a JSON schema, reproducible mock configuration, committed tests, and an internal benchmark report, giving the key behavior some traceability. Points are deducted because this review was static and did not execute anything; the benchmark covers only two tasks and provides no independently reproducible run records or corroborating evidence types, so a higher static score is not justified.
- Do not use live or mock output for diagnosis, treatment, triage, patient-facing reports, or other clinical decisions.
- Before using hf-space-api, confirm that patient images may be sent to an external Hugging Face service and assess mainland-China reachability, privacy, and compliance requirements.
- Prefer an isolated virtual environment or container for dependency installation; model weights and caches may consume substantial disk space.
- Mock mode verifies input handling, prompt wiring, and the JSON contract only; it does not validate model inference quality.
What does this skill do, and when should you use it?
This skill is a thin wrapper in NVIDIA’s skills collection for running NV-Reason-CXR-3B chest X-ray inference tests. It invokes either a local Hugging Face Transformers path or the public Hugging Face Space API and emits JSON. It accepts real images, JSON fixtures, and deterministic mock runs. It is explicitly not for diagnosis, clinical reporting, treatment decisions, triage, or clinical deployment.
Reads PNG, JPEG, or JSON fixture inputs; invokes scripts/run_nv_reason_cxr.py against the local model or Hugging Face Space API; routes prompts for analysis, report generation, or specific questions; can create a temporary synthetic fixture image; and emits JSON containing input metadata, prompt, model information, runtime mode, response text, and known limitations. Mock mode validates command shape and wrapper behavior.
- A developer needs CI coverage for the chest X-ray wrapper’s command arguments and JSON schema.
- A model engineer has CUDA and cached or downloadable model assets and wants live inference on a chest X-ray image.
- An integration engineer wants to test inference through the public Hugging Face Space without installing local PyTorch, Transformers, or CUDA.
- A test engineer needs a deterministic synthetic-image smoke test that does not download model weights.
- A researcher needs model context for analysis or structured-report drafting while treating the output as engineering evidence rather than a clinical conclusion.
What are this skill's strengths and limitations?
- Provides a fixed, reproducible script entrypoint and JSON output.
- Supports local Transformers inference, public API inference, and deterministic mock testing.
- Accepts PNG, JPEG, and generated JSON fixtures for automated validation.
- Documents prompt routing, backend selection, setup checks, and troubleshooting.
- Explicitly unsuitable for diagnosis, clinical interpretation, treatment decisions, triage, or regulatory use.
- Local live inference generally requires CUDA/GPU and may download model weights and remote code.
- The model may hallucinate, miss subtle abnormalities, misread support devices, or produce overconfident prose.
- The public API depends on Hugging Face Space availability and API compatibility.
- Model weights are subject to the NVIDIA OneWay Noncommercial License Agreement; the source does not state operating costs.
How do you install this skill?
Install the skill from the NVIDIA collection with: npx skills add nvidia/skills --skill nv-reason-cxr --yes. A client-specific example is: npx skills add nvidia/skills --skill nv-reason-cxr --agent codex --yes. The README does not document a separate clone or manual skill-folder installation flow. For local live inference, the documented dependencies are installed with: pip install torch==2.7.1 torchvision==0.22.1 transformers==4.56.1 Pillow.
How do you use this skill?
Run from the repository root. For a command-shape or fixture smoke test: python skills/nv-reason-cxr/scripts/run_nv_reason_cxr.py PATH_TO_CXR_OR_FIXTURE --mock --out-dir OUT_DIR. For local live inference: python skills/nv-reason-cxr/scripts/run_nv_reason_cxr.py PATH_TO_CXR_OR_FIXTURE --prompt "Find abnormalities and support devices." --backend local. For the public API, use --backend hf-space-api. Before downloading assets or running inference, use python skills/nv-reason-cxr/scripts/run_nv_reason_cxr.py --check-setup. Do not substitute undocumented commands such as eval_engine/run.py, infer.py, or python -m nv_reason_cxr for ordinary wrapper runs.
How does this skill compare with similar options?
The skill offers two documented inference routes—local Hugging Face Transformers and the public Hugging Face Space API. Mock mode is for wrapper and CI validation, not model inference.