gemma-4-e2b-ai-text-detector

gemma-4-E2B AI-text detector (on-device, GPU)

QAT-LoRA + linear-probe classifier on google/gemma-4-E2B-it-qat-mobile-transformers for AI-text detection. Quantization-aware trained (int8 activations + int8 KV cache) so runtime accuracy matches the unquantized model.

Files

• model.litertlm (2.2 GB) — base bundle: quantized weights equivalent to litert-community/gemma-4-E2B-it-litert-lm, int8 composite KV cache, static int8 activations, 200 LoRA input sockets, activations decode output.
• lora_adapter.tflite (11 MB) — trained detector LoRA (attention q/k/v/o, r=8).
• head.tflite (41 KB) — linear-probe classifier: [1,1536] activations -> 4 bucket logits.
• ARCHITECTURE.md — signature/socket spec + runtime integration notes.

Accuracy (true runtime: int8 activations + int8 KV cache, full test set)

• human-vs-AI: 0.986 (AUC 0.999) · human-vs-rest: 0.919 · ai-vs-rest: 0.891 · ternary: 0.821
• Per-source human-vs-AI: twitter 0.962, amazon 0.978, google_reviews 0.987, news 0.992, fineweb 0.998, reddit 0.997.
• Detects all providers (mean score): gemini 0.80, gpt 0.76, claude 0.72.

Deployment

• GPU backend (iOS Metal / ml_drift): verified, base model 33/35, ~4 s.
• CPU (XNNPACK) not supported (int2 MLP layers are GPU/NPU-only). Use GPU.
• Inference: clean-text preprocess (lowercase/normalize, NO chat template, no special tokens) -> prefill -> 1 decode -> read activations -> head.tflite -> softmax -> score = p·[0,1,2,3]/3. Calibrate the decision threshold on runtime scores.