The LoRA training targeted language/text modules; audio modules were not targeted, and vision fine-tuning was disabled. Reusing the official audio/vision components keeps the browser package compatible with the public Gemma 4 E2B ONNX runtime contract while using the trained Lisper text stack.

Evaluation

The release-quality evaluation result belongs to the full v18 hybrid acoustic+Gemma pipeline, not to a separate browser-only q4f16 eval:

Held-out rows: 2,000
Hard errors: 0
Verdict: pass
Class match: 0.976
Clear/non-clear match: 0.989
Exact four-line format: 1.0

This q4f16 package is the browser demo artifact for consumer-device testing.

App Use

Configure the app with:

bash

VITE_LISPER_BROWSER_MODEL_ID=thomasjvu/lisper-gemma4-e2b-audio-onnx-q4f16
VITE_LISPER_BROWSER_DTYPE=q4f16

Expected required browser payload is about 3.15 GB. Keep q4f16 as the primary browser dtype for the hackathon package.

Companion Artifacts

LoRA adapter: thomasjvu/lisper-gemma4-e2b-audio-lora
Merged full checkpoint: thomasjvu/lisper-gemma4-e2b-audio-full
Server-side ZeroGPU fallback: thomasjvu/lisper-zerogpu