Text Classification
ONNX
Laya
r

Laya: ONNX exports

ONNX exports of the Laya checkpoints by convaiinnovations, for use from R with layaR and from any other client of the laya-onnx Rust runtime. The weights are unchanged: this repository only converts them to ONNX. All credit for the models goes to the original authors.

This is an unofficial redistribution. It is not made, reviewed or endorsed by convaiinnovations, and "Laya" is used here only to name the models it contains.

Checkpoint Path in this repo Size
english / (root) 1.6 GB
multilingual multilingual/ 1.3 GB
typed-decisions typed-decisions/ 1.6 GB

The layout follows the upstream repository: English at the root, the other checkpoints in subfolders. Each folder holds model.onnx, model.onnx.data, rl_agent_config.json and tokenizer/.

Use from R

# install.packages("layaR", repos = c("https://gueyenono.r-universe.dev", getOption("repos")))
library(layaR)
laya_download_model("multilingual")

How the exports were made

Each export was made with upstream's laya-rust/tools/export_onnx.py --verify (opset 18, CPU), from the checkpoints in convaiinnovations/laya. Verification compares ONNX Runtime with PyTorch on recorded inputs and on a sweep of 36 input shapes. The largest difference in logits was below 1e-5 for every checkpoint.

Changes from the original

The PyTorch checkpoints were converted to ONNX (model.onnx plus external weights in model.onnx.data). The weights, configuration (rl_agent_config.json) and tokenizers are unchanged. Nothing was retrained or quantised.

License and credits

  • Laya checkpoints: Apache-2.0, by convaiinnovations. This repository is distributed under the same license; see LICENSE.
  • Base encoders: the english and typed-decisions checkpoints are fine-tuned from ModernBERT-large (Apache-2.0, Answer.AI and LightOn). multilingual is fine-tuned from mmBERT-base (MIT, JHU CLSP). Their notices are in THIRD_PARTY_NOTICES.md.

See the upstream model card for the models' intended use, training data and limitations. If you use the models in research, cite the original work:

M, N.; M, A. From Laya to Vega: Open Non-Autoregressive Decision Models with Calibrated Probabilities. Preprints 2026, 2026110781.

and the base encoders:

@misc{modernbert,
  title  = {Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory
            Efficient, and Long Context Finetuning and Inference},
  author = {Benjamin Warner and Antoine Chaffin and Benjamin Clavi\'e and Orion Weller and
            Oskar Hallstr\"om and Said Taghadouini and Alexis Gallagher and Raja Biswas and
            Faisal Ladhak and Tom Aarsen and Nathan Cooper and Griffin Adams and Jeremy Howard
            and Iacopo Poli},
  year   = {2024}, eprint = {2412.13663}, archivePrefix = {arXiv}, primaryClass = {cs.CL},
  url    = {https://arxiv.org/abs/2412.13663}
}

@misc{marone2025mmbertmodernmultilingualencoder,
  title  = {mmBERT: A Modern Multilingual Encoder with Annealed Language Learning},
  author = {Marc Marone and Orion Weller and William Fleshman and Eugene Yang and Dawn Lawrie
            and Benjamin Van Durme},
  year   = {2025}, eprint = {2509.06888}, archivePrefix = {arXiv}, primaryClass = {cs.CL},
  url    = {https://arxiv.org/abs/2509.06888}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for gueyenono/laya-onnx

Quantized
(94)
this model

Papers for gueyenono/laya-onnx