---
license: mit
task_categories:
  - image-to-text
  - visual-question-answering
tags:
  - vlm
  - document-intelligence
  - quantization
  - benchmark
  - synthetic
---

# VisionFlow evaluation set and benchmark results

Companion data for [VisionFlow](https://github.com/vineetha00/visionflow), an
edge-first quantized VLM pipeline for local document intelligence.

## Contents

| Path | What it is |
|---|---|
| `sample_images/` | Three synthetic document images: a surgical report, a cold-chain shipping manifest, and a bar chart |
| `labeled_set.json` | Ground-truth field labels for those images |
| `results/` | Raw output from `vf bench`, `vf accuracy`, and the ONNX provider benchmark |

## Ground truth

Labels are the literal strings drawn into each image by the generator script, so
they are exact by construction rather than human-annotated. The trade-off is scope:
this is 3 images and 21 fields, sized to expose *relative* differences between
quantization levels — not to support an absolute accuracy claim.

## Synthetic data notice

Every name, MRN, date, diagnosis code, SKU, lot number, and revenue figure here is
fabricated. No real patient, shipment, or company data is present. The medical
image is a plausible-looking surgical report and is not a real medical record.

## Reproducing

```bash
pip install visionflow
vf bench
vf accuracy
```