SuperPrompts AI text detector (ONNX, int8)

Browser-ready ONNX export of fakespot-ai/roberta-base-ai-text-detection-v1 (Apache-2.0, by Mozilla's Fakespot team), dynamically quantized to int8 (≈125 MB). Used by SuperPrompts to detect AI-written text entirely on the user's device with onnxruntime-web.

  • model.onnx — inputs input_ids, attention_mask (int64, max 512 tokens); output logits [human, ai]
  • vocab.json, merges.txt — RoBERTa byte-level BPE tokenizer
  • Labels: 0 = human, 1 = ai

Evaluation (91 texts: 50 human + 20 ChatGPT from HC3, 20 modern gpt-oss, 1 QuillBot AI Chat)

AUROC AI caught (≥0.5) Human flagged (≥0.5)
fp32 original 0.993 41/41 7/50
int8 (this repo) 0.991 40/41 7/50

Like all AI-text detectors it can misjudge some human writing (e.g. famous classic literature). Use scores as a signal, not proof.

All credit for the model goes to the original authors; this repository only changes the file format.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for backuplan/superprompts-ai-detector

Quantized
(6)
this model