Gemma 3 1B IT Brand Advisor (merged, ONNX quantized)
LoRA-finetuned google/gemma-3-1b-it for automotive recommendations. Merged adapter into base, exported to ONNX, then dynamically quantized (INT8, arm64) for browser/server inference.
- Base:
google/gemma-3-1b-it
- Task: Single-paragraph car recommendation from visitor info.
- Artifacts:
model_quantized.onnx (INT8, ~418 MB), tokenizer and config from the same merged model for correct decoding.
Note: Tokenizer and config in this repo are from the same merged_model used for ONNX export to avoid vocab mismatch. True 4-bit ONNX would require a separate toolchain (e.g. onnxruntime-genai); this upload uses INT8 quantization.
