anusreevjn/ai-detector-onnx

AI text detector (ONNX, browser ready)

A browser-ready ONNX conversion of desklib/ai-text-detector-v1.01, a DeBERTa-v3-large model that classifies English text as human-written or AI-generated.

What changed from the original

• Exported to ONNX (opset 17), including the mean pooling and the single-logit classifier, so the graph takes input_ids and attention_mask (int64, shape [batch, sequence]) and returns logits (shape [batch, 1]). Apply a sigmoid to get the probability that the text is AI-generated.
• Matrix multiply weights stored as 8-bit (weight-only, block size 32), word embeddings stored as 8-bit, and precomputed relative position projections stored as 16-bit. Activations stay in 32-bit float. Size went from 1.74 GB to 524 MB.
• tokenizer.json and tokenizer_config.json are the original files from the source repository.
• The model is split into model.onnx.part1 and model.onnx.part2 so each file stays small enough to share easily. Join them in order to get model.onnx (SHA-256 18486ab6ecc62e67c2fb98b1f03928a2bbf68e5d553d7259e3c39e8c0d831b9d). The AI Detector website does this in the browser automatically.

Accuracy check

Compared against the original PyTorch model on 24 texts from 10 to 768 tokens, the largest probability difference was 0.0087 (under 1 percentage point). Running in onnxruntime-web 1.30.0 (WebAssembly) in Chromium gave the same scores within 1 percentage point.

Dynamic 8-bit quantization of activations was tested and rejected: it raised scores on human-written text by up to 80 points.

Use

Truncate to 768 tokens, keeping the final [SEP] token (id 2). The original model card recommends a 0.5 threshold.

License

MIT, same as the original model. All credit for training goes to Desklib.