Initial synthetic receipt calibration set

256 synthetic PNG receipts for calibrating both the Digit and VQA models. These are the exact initial synthetic 256 images used in the 2026-10-27 comparison, copied without resizing, re-encoding or selecting different samples. They were produced by the untrained initial rendering policy, before any RL updates. Every image includes a watermark.

The images occupy 61,200,033 bytes (58.36 MiB). Paths in manifest.json are relative to this directory; its record order is 000000 through 000255. Each record contains the original PNG hash, decoded RGB hash and dimensions, plus the path and hash of its matching synthetic annotation JSON. The 256 annotation files occupy another 5,721,582 bytes (5.46 MiB). Post-training calibration uses image pixels and model-generated decoder inputs; it does not use the supplied answer labels.

Digit QAT uses numeric labels from the annotations; VQA PTQ uses image pixels and generated decoder inputs. The default split and recipes are documented in Build. See the benchmark report for recorded synthetic/real calibration comparisons and their evaluation limits. This calibration set is not an independent accuracy benchmark and does not qualify a fresh model build.

Receipt contents and annotations

Each image has a matching JSON: images/000000.png pairs with annotations/000000.json. These are unchanged generated labels from the original synthetic receipts. Their two top-level sections are:

answers: receipt contents, including store, address, street_no, phone_disp, phone_digits and items. Each item has name, quantity, unit_price and total. It also contains language and layout metadata. Some receipts have no store or item list; check has_store and has_items.
fields: the rendered text fields and their polygons, including glyphs[].char and glyphs[].poly for individual characters.

For example, the content in annotations/000000.json includes:

{
  "store": "싸요 클럽쉼터 직영점",
  "address": "오산시 예천군 옛군청옆길 868",
  "street_no": 868,
  "phone_disp": "161-1627",
  "phone_digits": "1611627",
  "has_items": false,
  "items": []
}

Coordinates refer to the original 670 × 320 PNG pixels. If you resize an image to the models' 672 × 320 input, scale x coordinates by 672 / 670 when using these polygons. Added watermark text is not part of fields. The renderer changes appearance while preserving the receipt's geometry and label bytes.

Use with this repository

From the repository root, first verify the downloaded files:

(cd synth && sha256sum --check --quiet SHA256SUMS)

Use the main build instructions for default training, calibration and cache reuse. Optional precision profiles and custom calibration inputs are documented with each model:

For another quantizer, load the PNGs in manifest order and use the preprocessing below. The VQA encoder also needs tokenized questions, attention masks and its family-routing inputs; its decoder needs valid generated token/KV trajectories. Images alone are not a complete VQA calibration feed. The original experiment's two questions per image and decoder-prefix settings are recorded under reference_quantization in the manifest.

Preprocessing

Keep the distributed PNG files unchanged. At input time, both models use:

Convert to grayscale with Pillow Image.convert("L").
Resize to width 672, height 320 with Image.Resampling.BILINEAR.
Convert to float32 and normalize with (pixel / 255.0 - 0.5) / 0.5.
Arrange as NCHW; one image has shape [1, 1, 320, 672].

The repository build applies this preprocessing automatically. The PNGs contain their original RGB pixels and original dimensions, which can differ from the model input size.

Origin and scope

Paper/background pixels and font glyphs were rendered procedurally; captured receipt images and crops were not composited into these images. Store names, address prefixes and item names were sampled from a text bank built from existing receipt annotations. Phone digits and street numbers were generated randomly. The images therefore should not be described as using wholly invented text. Watermarks were rendered with Noto Sans CJK Regular; font binaries are not bundled in this dataset.

manifest.json records the generation stage, seed, source artifact hashes and pinned FP32 model identities. The public files contain portable paths and include only the matching generated synthetic annotations. Captured-receipt annotation files, private evaluation data and experiment logs are excluded.

This is a calibration set, not an independent accuracy benchmark. The original comparison used ONNX Runtime 1.23.2 static U8S8 QDQ; this repository's commands use VolvoxAI with different conversion, question and decoder-prefix settings. Sharing the same images does not imply identical quantized weights or the same accuracy. The original comparison covered phone/address questions and does not establish performance for all VQA task families.

License

The distributed synthetic images, generated annotations and dataset documentation/metadata are provided under the repository's Apache-2.0 license. See NOTICE for project and third-party attribution.