Finalize Lisper E2B full model card

README.md

@@ -7,32 +7,49 @@ tags:
 - audio
 - merged
 - unsloth
+- speech
 - lisper
 ---
 
 # Lisper Gemma 4 E2B Audio Full Checkpoint
 
-This is the v16 merged Lisper checkpoint: `google/gemma-4-E2B-it` with the trained Lisper LoRA adapter folded into a standalone `safetensors` model.
+This is the merged Lisper checkpoint: `google/gemma-4-E2B-it` with the trained Lisper LoRA adapter folded into a standalone `safetensors` model.
 
-## Validation
+## Model Lineage
 
-- Source run: Kaggle notebook `thomasjvu/lisper-gemma-4-audio-unsloth-training`, v16
-- Merge method: `merged_16bit`
+- Base model: `google/gemma-4-E2B-it`
+- Training: Unsloth supervised fine-tuning with QLoRA / LoRA
+- Merge method: base + LoRA adapter merged into a 16-bit checkpoint
 - Weight file: `model.safetensors`
-- Safetensors validation: complete
-- Size: `10,246,621,886` bytes
-- Tensor count: `2011`
+- Training rows: `16,000`
+- Validation rows: `2,000`
+- Held-out test rows: `2,000`
 
-## Training Summary
+This is not a dense full-parameter fine-tune. It is a merged base+LoRA checkpoint for easier deployment.
 
-- Method: supervised fine-tuning with Unsloth QLoRA / LoRA, then merged export
-- Training steps: `4000`
-- Selected checkpoint: `checkpoint-2500`
-- Final training loss: `0.7217483117580413`
-- Best validation metric: `4.851842880249023`
+## Evaluation
 
-This artifact is the canonical full-model checkpoint for export and evaluation. It is not a dense full-parameter fine-tune; it is a merged base+LoRA checkpoint.
+The release-quality evaluation is the v18 hybrid acoustic+Gemma path:
 
-## Deployment Notes
+- 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`
+- Reason/cue/encouragement present: `1.0`
 
-Use this checkpoint as the correctness reference. For browser and consumer-device deployment, derive a 4-bit ONNX/WebGPU artifact and evaluate that quantized runtime before wiring it into the app.
+The evaluated pipeline uses acoustic features for the lisp-class hint and Gemma for structured coaching text and tone. Do not interpret these metrics as a pure direct-Gemma raw-audio classification result.
+
+## Deployment
+
+- Browser/WebGPU demo: `thomasjvu/lisper-gemma4-e2b-audio-onnx-q4f16`
+- Server-side ZeroGPU fallback: `thomasjvu/lisper-zerogpu`
+- Adapter-only package: `thomasjvu/lisper-gemma4-e2b-audio-lora`
+
+Use the q4f16 ONNX/WebGPU package for browser demos. Use this merged checkpoint as the server-side correctness reference.
+
+## Limitations
+
+- The lisp dataset is synthetically generated from speaker-disjoint source speech.
+- This is a practice assistant, not a medical diagnosis tool or a replacement for a speech-language pathologist.