ternlang-core 0.3.3

Compiler and VM for Ternlang — balanced ternary language with affirm/tend/reject trit semantics, @sparseskip codegen, and BET bytecode execution.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

// Module:  stdlib/nlp/positional.tern
// Purpose: Positional Encoding for NLP
// Author:  RFI-IRFOS
// Ref:     https://ternlang.com

// Absolute and relative positional encodings.

fn sinusoidal_trit(pos: int, dim: int) -> trit {
    let cycle: int = pos % 3;
    if cycle == 0 { return affirm; }
    if cycle == 1 { return tend; }
    return reject;
}

fn rotary_pe_trit(feature: trit, pos: int) -> trit {
    // RoPE analog. Rotates state based on position.
    if pos % 2 == 0 {
        if feature == affirm { return tend; }
    }
    return feature;
}

fn alibi_trit(distance: int) -> trit {
    // ALiBi adds a linear bias based on distance
    if distance > 10 { return reject; } // Far away, reduce attention
    if distance == 0 { return affirm; } // Self, high attention
    return tend; // Neutral
}