← Back to Live Labs

ARTI / DYNAMIC VOCABULARY LAB

The input and output alphabets do not have to match.

The model reads glyph tensors, selects local slots from an output vocabulary supplied for this call, and composes them back into the visible input. Diacritic recognition is only one part of reconstruction.

TRAINED WEIGHTS · DYNAMIC OUTPUT VOCAB

Give the model a new alphabet at runtime.

The input side receives visible glyph tensors only. The output side is an independent vocabulary supplied for this call. The model selects local base-glyph and combining-mark slots; a deterministic compositor reconstructs the input.

REAL ONNXARTI 1.3.0
WASM VERIFIEDGLYPH → LOCAL SLOTSDynamic K=8–14Loading the browser model…
Try

glyphs [B,1,32,32] · pixels only · no code point/token IDs
02 / Edit this call’s output vocabularyoutput_vocab [B,K,1,20,20]

Slot numbers exist only for this call. Remove a required base or mark to inspect REJECT behavior.

K = 0
    03 / Reconstructionbase + mark local slots
    Inputnaïve
    →
    Output—
    ○
    Exact visible reconstruction—Pixel + string agreement: —
    Local slot path
      98.93%Target present
      90.04%Target absent
      98.83%Vocabulary permutation
      93.80–96.29%Dynamic K=8–14

      Controlled closed-set Latin grapheme reconstruction—not general or open-vocabulary OCR. The model receives visible pixels only; a zero-width code point leaves no visual evidence and cannot be recovered from a glyph tensor.

      Engineering evidence
      Input contract
      visible glyph tensors · no character or token IDs
      Output contract
      independent per-call vocabulary · two local slots per grapheme
      Model
      trained ARTI RuntimeVocabEncoder + LiteralOutputHead
      Parity
      PyTorch → ONNX → WASM · opset 18
      Browser runtime
      Exported by arti-fit 1.3.0 · executed by @arti-fit/web 0.1.0-alpha.3
      WebGPU parity
      Not run