MELD AI text detector (ONNX, browser ready)

A browser-ready ONNX conversion of the MELD AI-generated text detector (odeveloper/meld), a ModernBERT-based model that leads the open-source RAID benchmark and is strong at low false-positive rates.

What this is

meld_int8.onnx plus meld_int8.data: the ModernBERT backbone exported to ONNX and 8-bit weight-only quantized (block size 32). The matrix-multiply weights are stored as 8-bit, cutting the model from 4.1 GB to 1.4 GB.
meld_int8.data is split into meld_int8.data.part0 through part3 so each file stays easy to download. Join them in order to get meld_int8.data (SHA-256 6f408c36aee767e0d07e4d547737eec7b60d8e7c463a72b754a1f6c62c1f53cd). The AI Detector website does this in the browser automatically.
meld_head.json: MELD's metric-learning head (style projection, human anchors, family prototypes, pooling settings). The website reimplements the head in JavaScript and runs it on the backbone output, so the ONNX file only needs to hold the backbone.
tokenizer.json and tokenizer_config.json: the original ModernBERT tokenizer from the source model.

Accuracy

The ONNX backbone plus the JavaScript head reproduce the original MELD scores within 0.5 percentage points, in Python and in onnxruntime-web (WebAssembly).

How the score is formed

Each window of up to 2046 tokens is run through the backbone. The head turns each token's hidden state into a score against learned human anchors and AI family prototypes, and the document score is the mean of the top-scoring tokens, passed through a sigmoid. See meld.py in the source repository for the reference implementation.

License

MIT, same as the source model. All credit for training goes to the MELD authors.

Downloads last month: -; Downloads are not tracked for this model. How to track