GabeuxDev/ai-text-detector-v1.01-onnx

ai-text-detector-v1.01 — ONNX (fp32)

Unmodified ONNX export of desklib/ai-text-detector-v1.01 (MIT license), a DeBERTa-v3-large based AI-generated-text detector. All credit for the model and its training belongs to the original authors; this repository only converts it so it can run with onnxruntime and Transformers.js (e.g. fully in the browser).

Details

• —Graph: mean pooling + single-logit classification head baked into the ONNX graph. Inputs input_ids, attention_mask; output logits of shape [batch, 1]. Apply a sigmoid to get P(AI-generated). config.json sets problem_type: multi_label_classification so Transformers.js applies the sigmoid automatically (single label: AI).
• —External data format: the graph lives in onnx/model.onnx (~2 MB) and the fp32 weights in onnx/model.onnx_data (~1.7 GB). Native onnxruntime picks the data file up automatically; in Transformers.js pass use_external_data_format: true. This format is required in browsers — a single-file 1.7 GB model gets double-buffered during parsing and exceeds the 4 GB WebAssembly memory limit. Even with external data, in-browser inference at this size needs WebGPU (device: "webgpu"); the CPU/WASM backend cannot hold the weights plus activations.
• —Precision: fp32 only, on purpose. int8 quantization shifts this model's probabilities by up to 0.5 (DeBERTa's disentangled attention is quantization- sensitive), and fp16 is pathologically slow on onnxruntime CPU. fp32 matches the original PyTorch outputs exactly.
• —Sequence length: exported for up to 768 tokens (relative position attention).

Usage (Transformers.js)

import { pipeline } from "@huggingface/transformers";

// Requires transformers.js v4+. v3.x mis-tokenizes DeBERTa prefix spaces
// (first-word token differs from the Python tokenizer) and produces NaN on
// padded WebGPU batches; both are fixed in 4.x (verified on 4.2.0).
const detect = await pipeline(
  "text-classification",
  "GabeuxDev/ai-text-detector-v1.01-onnx",
  { dtype: "fp32", device: "webgpu" }
); // config.json already sets use_external_data_format for the loader
const [{ score }] = await detect("Text to check.");
console.log("P(AI) =", score);

Note: the model file is ~1.7 GB; in a browser this implies a large one-time download and several GB of RAM during inference.

License

MIT, inherited from the original model. See the original repository for model details, training information and evaluation.