Gemma 4 E4B for A2UI Express (retrain 2026-11-07)
A browser-ready Gemma 4 E4B fine-tune that generates the A2UI Express DSL for
the CopilotKit Express demo: the 160-step Unsloth LoRA run on 240 training
conversations (the original 159 form rows plus 90 informational cards), adapter
SHA-256 778467f51be80c125d41cecfc654f20083013e8602926f2ee1b7d7de910e105b,
validation loss 0.038. The E2B companion,
soverius-ai/gemma-4-E2B-it-A2UI-Express-ONNX, was trained on the larger
multi-turn set and is the better-tested of the two.
Files
onnx/: theonnx-community/gemma-4-E4B-it-ONNXq4f16 text sessions (graph revision2b6d2e26a962f70a33da8f7cd81e1afdd8d9cfb2) with the 300 adaptedMatMulNBitsmatrices requantized from the merged weights (4-bit, block size 32, asymmetric, ONNX Runtime quantizer). The decoder graph and the embedding data are byte-identical to the base export.a2ui-export-manifest.json: the provenance descriptor the demo checks.adapter-mlx/: the LoRA adapter (mlx_lmlayout, rank 8, scale 2.0, 342 logical matrices).system_prompt.txt: the system prompt the demo sends and the adapter was trained with.
Use
Load with @huggingface/transformers (dtype: "q4f16", device: "webgpu",
shader-f16 required), send system_prompt.txt as the system message, and
validate the output with the A2UI Express compiler. Gemma weights keep
Google's Gemma terms; see NOTICE.txt.