Improved multi-dimensional array representation

[debonet]

Instead of:

// Array of arrays
{ pairs: [[1, 2], [3, 4]] }     → pairs[2]:
                                    - [2]: 1,2
                                    - [2]: 3,4
                                    

How about

// Array of arrays
{ pairs: [[1, 2], [3, 4]] }     → pairs[2][2]:
                                   - 1,2
                                   - 3,4
                                   

?

This would make representation of multi-dimensional datasets far more efficient. E.g. tensor[3][3][3]:

[johannschopplich]

Tempting!

The list format is intentional: each - [2]: acts as an explicit length marker for that row. When LLMs parse pairs[2][2]:, they need to infer the inner structure from delimiters alone. Current format prioritizes readability; models see the shape per-row instead of reconstructing it from context. The current TOON spec trades tokens for clarity/familiarity.