---
license: cc0-1.0
base_model: Qwen/Qwen2.5-Coder-3B-Instruct
library_name: peft
tags: [lora, emergent-language, agent-communication, code, multi-task]
---
# Glyph v2 — multi-task LoRA (speaker · builder · translator in one adapter)

Finetuned on 40,592 pairs over a 68-primitive typed glyph language
(dataset: robertkeus/glyph-tasks, sft2.jsonl). One adapter, three roles by
instruction prompt:

- `Encode this task as glyph symbols.\nTask: {english}\nSymbols:`
- `Write Python for this glyph message.\nSymbols: {glyphs}\nCode:`
- `Translate this glyph message into English.\nSymbols: {glyphs}\nEnglish:`

## Results (seed 0, 3 epochs, Kaggle T4, LoRA r=32 all-linear)
| metric | score |
|---|---|
| builder: glyphs → correct code, HELD-OUT compositions (tests executed) | **99.5%** |
| speaker: UNSEEN phrasings → exact glyph message | **86.5%** |
| zero-shot symbols (never trained) | 2% |

Code + demo: https://github.com/robertkeus/glyph-ai