soverius-ai/gemma-4-E4B-it-A2UI-Express-LoRA

Gemma 4 E4B LoRA for A2UI Express (retrain 2026-11-07)

An mlx_lm-layout LoRA adapter (rank 8, scale 2.0, 342 logical matrices, SHA-256 778467f51be80c125d41cecfc654f20083013e8602926f2ee1b7d7de910e105b) trained with Unsloth for 160 steps on 240 conversations (the original 159 form rows plus 90 informational cards) with the CopilotKit Express demo's system prompt. Base: unsloth/gemma-4-E4B-it at 4e22d7e59e078e63a14f351efdc5232ed366b621; validation loss 0.038. Offline it compiles 3 of the demo's 4 example prompts; it has not been exported for the browser yet. Gemma weights keep Google's Gemma terms; see NOTICE.txt.