After seeing so many enthusiastic posts about TOON (Token-Oriented Object Notation), the bold new data format that came to replace JSON - I thought of something even better. 😁
This should be a truly visionary format: VSBC - Values Separated By 🥕.
Yes, the carrot. The nature's comma. 😜
Instead of:
products[3](id,name,category,price):
1,"Laptop","Computers",399
1,"Mouse","Peripherals",14
1,"Headset","Accessories",49
Behold:
1🥕Laptop🥕Computers🥕399
1🥕Mouse🥕Peripherals🥕14
1🥕Headset🥕Accessories🥕49
No braces, quotes, escapes or syntactic noise, and no fear that one rogue comma or double-quote detonates the entire file.
Just pure, uncut data separated by a root vegetable.
The 🥕 is U+1F955 - that’s 4 bytes in UTF-8 but a single token for a LLM.
Yes, it’s longer than a comma (1 byte), but it removes the need for quoting or escaping overhead.
But if someone insists in saving more bytes, good UTF-8 alternatives are:
✦ (star) 3 bytes
¤ (generic currency sign) 2 bytes
^ (caret) 1 byte (my favorite delimiter when creating composite keys for maps)
In my very short experience with VSBC, adding 🥕 as a delimiter tended to produce shorter LLM responses.
I can only especulate - someone could do an investigation on that:
- could it be that 🥕 acts as a compression cue to the LLM that the user prefers
minimalism?
- could it be that emoji delimiters suppress narrative mode so the LLM tend to respond with less prose?
- do low-frequency patterns cause the model to output less?
- or does 🥕 give the LLM a hint of needing to go on a verbal diet?
😜
