unaveraged Digit

Reads Korean and English questions about phone-number digits and numeric street addresses. The default model is conv4-linear2 QAT, with 4-bit convolution and 2-bit grouped linear weights.

Use the main README for building, Chrome and Firefox installation, and testing existing packages, selecting MODEL=digit. The benchmark report contains model comparisons, measurement conditions and the tradeoffs behind the default profile.

Model and runtime

Images become grayscale float32 tensors of shape [1, 1, 320, 672], normalized to [-1, 1]. One forward pass produces 16 digit slots: 12 for a phone number and four for a street number. The adapter selects the digits requested by the question. A new question about a cached image can reuse the numeric record.

The default QAT recipe uses the bundled synthetic receipts and numeric annotations. The exact training split, optimizer settings and backend options are documented in Build. Fresh builds can produce different weights and accuracy from the recorded benchmark artifacts.

The public FP32 source is tiny-receipt-reader-digit-slots-2m at revision c9fa07886f99334ea820bfe312d6b4771d9460b0. Optional artifact imports verify the exact files and provenance against frozen-profiles.json. Fresh builds do not have to match those historical artifact hashes.

Optional fresh W8A8 build

W8A8 converts the pinned public FP32 source and calibrates its 11 convolutions with INT8 weights and activations; dense layers remain FP32. From the repository root, use the bundled calibration images:

make MODEL=digit PRECISION=w8a8 CALIBRATION_IMAGES=synth/images

For a separate calibration set, exclude external evaluation images explicitly:

make MODEL=digit PRECISION=w8a8 CALIBRATION_IMAGES=/path/to/calibration/images \
  EVAL_IMAGES=/path/to/evaluation/images

This writes dist/digit/w8a8/. Provide 256 distinct JPG/JPEG/PNG/WebP calibration images directly inside the directory, or set CALIBRATION_COUNT. Public examples and EVAL_IMAGES are excluded from calibration. The default seed is 20261027. Without EVAL_IMAGES, no external evaluation-overlap check is recorded.

Use the test workflow with MODEL=digit PRECISION=w8a8. Optional external evaluation is available through .venv/bin/python digit/check.py --evaluation-root /path/to/evaluation, with images/ and annotations/ subdirectories. External data and detailed reports stay outside extension packages. Data handling is covered by the privacy policy.