| license: cc0-1.0 | |
| base_model: Qwen/Qwen2.5-Coder-3B-Instruct | |
| library_name: peft | |
| tags: [lora, emergent-language, agent-communication, code] | |
| # Glyph — trained LoRA adapters | |
| Three LoRA adapters on Qwen2.5-Coder-3B for a seeded logographic agent-language: | |
| - **speaker**: English task → glyph symbols | |
| - **builder**: glyph symbols → Python code | |
| - **translator**: glyph symbols → English | |
| Results (seed=0): compositional decode 12/22, cold-model self-explanation 13/22, | |
| byte reduction 98.8% vs English agent-to-agent. Code + demo: | |
| https://github.com/robertkeus/glyph-ai | |