base64

text in, base64 out · 0 bytes ·


every character above was emitted by ollama, one token at a time. the page sends bytes and concatenates tokens — it never encodes anything

server log — 0 messages 





  
A base64 encoder in the weights

  
An 18 MB GGUF that encodes base64, served by a stock ollama. The
     page sends raw bytes as <B:xx> tokens and reads the
     characters back — it never encodes anything itself.

  
Two transformer blocks: the first works out where to read — each
     output character depends on which bytes sit at fixed offsets from the
     end marker — and the second reads them and applies the base64 rules
     (6-bit slicing and padding, one MLP gate per case). No training,
     every weight constructed, verified byte-for-byte against Python's
     base64 for every input length.

  
Inputs are chunked 21 bytes per call: base64 of a multiple of 3
     concatenates cleanly, and 21 bytes can emit at most 28 identical
     characters in a row — just under ollama's anti-loop guard, which cuts
     generation at ~31 identical tokens (try the 24 zero bytes
     button to see why this matters).