SciNCL, INT8 ONNX
A dynamically quantized ONNX export of malteos/scincl,
for CPU inference through onnxruntime. Nothing here is a new model: the weights
are SciNCL's, converted and quantized.
| file | what it is |
|---|---|
model_quantized.onnx |
the SciNCL encoder, INT8 dynamic quantization (weights QInt8); outputs token embeddings |
tokenizer.json |
SciNCL's tokenizer |
To get SciNCL's document embedding, take the [CLS] token's vector (768 dimensions), as the original model's pooling configuration does.
Quantization changes the numbers. These are INT8 weights, not the original float weights, so embeddings and scores differ slightly from the original model. Do not treat this as a drop-in reproduction of published SciNCL results.
Credit and licence
SciNCL is by Malte Ostendorff and co-authors:
Ostendorff, Rethmeier, Augenstein, Gipp, Rehm. Neighborhood Contrastive Learning for Scientific Document Representations with Citation Embeddings. EMNLP 2022. arXiv:2202.06671
The original model is released under the MIT licence, Copyright (c) 2022 Malte
Ostendorff. That licence and notice are in LICENSE and apply to
this conversion too.
This conversion is used by PaperReap (paperreap.com), which downloads it only when a user chooses SciNCL.
Model tree for Vranesh/scincl-onnx
Base model
malteos/scincl