Replaces the v2 network on main. Not a drop-in: input height 128 -> 160, output classes 316 -> 277, charset 315 -> 276, and the width axis goes from dynamic to a static 1024 because attention fixes the sequence length at trace time. Revision a51be11 still serves v2 and is unaffected.
Both exports are gated against the PyTorch model on a seeded uniform [-1,1] input: ONNX on logits at rtol=atol=1e-3, CoreML on logits at 1e-3 and on an identical decoded string. The CoreML gate runs CPU-only, so on-device numerics are unverified.
No held-out evaluation exists for this model. The card states what the 0.0210 selection metric is and is not.
fix: the sidecars declared a 912-character charset for a 316-class model
Both monocr.json and onnx/monocr.json carried the 912-character corpus-scanned
string. The model has 316 output classes, so exactly 150 of 315 indices mapped
to the wrong character and any consumer reading these files decoded Mon into
Latin and IPA with no error anywhere.
The correct 315-character string was already in this repository, in
charset.txt beside them. It was taken programmatically from that file, not
retyped, and it is not a truncation of the 912: the long string is missing
U+2026, which the real charset carries, so no consumer could have recovered by
trimming.
How it happened, because a re-export will do it again: df456df was a Core ML
export that wrote the CoreML sidecar over both other formats. That is also why
onnx/monocr.json said "format": "mlpackage" and had lost its opset_version.
The Core ML package still has no sidecar of its own.
- charset 912 -> 315 in both; onnx/ regains format "onnx" and opset_version 17
- root sidecar sits beside monocr.ckpt, so its format…