aiflow-math-ink-1.0e

Online handwritten mathematics recognition

Concept

This checkpoint represents mathematical handwriting as an ordered pen-trajectory sequence. It preserves stroke order and temporal structure, converts the sequence into compact memory tokens, and maps those tokens to a free-form mathematical text decoder.

The raster encoder is used only as an offline teacher while fitting the bridge. It is not part of the runtime path.

Functions

• Read ordered online ink directly without rasterization.
• Encode trajectory shape and timing through the online prior.
• Compress the sequence into 16 temporal memory tokens.
• Project the memory into the decoder cross-attention space.
• Produce free-form LaTeX mathematical text.
• Use a raster-derived latent target during fitting while keeping raster-free runtime inference.

Input and output

• Input: ordered raw strokes, up to 128 events, 19 trajectory channels.
• Output: free-form LaTeX text.
• Checkpoint: serial_bridge.pt contains the online adapter, sequence bridge, and tuned decoder cross-attention weights.
• Runtime assembly: the matching decoder base and the checkpoint are loaded together.