EditLens RoBERTa Model Kit — CoderBak

This repository redistributes the original Pangram EditLens RoBERTa-large
checkpoint and converted ONNX artifacts derived from that checkpoint.
It is a community conversion/distribution project, not a newly trained model.
No affiliation with or endorsement by Pangram, the EditLens authors, Meta,
Hugging Face, or Microsoft is claimed.

Original model developer: Pangram.
Original research authors: Katherine Thai, Bradley Emi, Elyas Masrour,
and Mohit Iyyer.
Paper: EditLens: Quantifying the Extent of AI Editing in Text.
https://arxiv.org/abs/2510.03154
Upstream model: https://huggingface.co/pangram/editlens_roberta-large
Pinned source revision: f93e1ace74528cfb48f337ab2fe946fb71a728cb
Upstream research code: https://github.com/pangramlabs/EditLens
Base-model lineage: https://huggingface.co/FacebookAI/roberta-large
The base-model card identifies its license as MIT. The EditLens derivative
weights distributed here retain Pangram's CC BY-NC-SA 4.0 license.

License: Creative Commons Attribution-NonCommercial-ShareAlike 4.0
International (CC BY-NC-SA 4.0).
https://creativecommons.org/licenses/by-nc-sa/4.0/
The complete license text is provided in LICENSE. Its attribution,
noncommercial, share-alike, and other applicable terms continue to apply.
Public availability does not grant commercial-use rights.
No additional research-only restriction is imposed by this model kit.
Third-party software dependencies retain their own licenses.

Changes made by CoderBak:
- Exported the complete sequence classifier to ONNX with opset 17,
  dynamic batch/sequence axes, int64 input IDs and attention masks,
  and four output logits.
- Created optional FP16 and dynamic per-channel INT8 MatMul variants.
- Added conversion scripts, local inference example, provenance,
  checksums, and numerical conversion checks.
- No retraining, distillation, new calibration, or change of label order.
- The root PyTorch checkpoint/configuration/tokenizer files are preserved
  byte-for-byte from the pinned upstream snapshot.

The original model card is retained in upstream/README.md. Its access-form
metadata is historical upstream information, not a gate for this repository.
This repository is public and ungated; that does not waive the model license
or grant access to the separately gated upstream repository.

The license includes a disclaimer of warranties and limitation of liability.
Please retain this notice, source attribution, license, and change notices
when redistributing the model or further derivatives.