# Attribution and model provenance

This is a mechanically converted, dynamically quantized derivative of:

- Original checkpoint: `almanach/xlmr-chatgptdetect-noisy`, by ALMAnaCH (Inria).
- Exact revision: `e7bc8107ddd1486b43cc9fe76ca647e70fb365f0`.
- Source: https://huggingface.co/almanach/xlmr-chatgptdetect-noisy
- Source model card declares `license: mit`. The original model card is retained as `MODEL_CARD_ORIGINAL.md`.
- Base architecture/checkpoint: XLM-RoBERTa base, Facebook AI. Base software MIT notice retained in `LICENSE-base-MIT.txt`.
- Research: Wissam Antoun, Virginie Mouilleron, Benoît Sagot, Djamé Seddah (2023), *Towards a Robust Detection of Language Model-Generated Text: Is ChatGPT that easy to detect?* https://aclanthology.org/2023.jeptalnrecital-long.2/
- Research source repository: https://gitlab.inria.fr/wantoun/robust-chatgpt-detection

The source model card's explicit MIT declaration is the fine-tuned checkpoint's license evidence. No separate fine-tune copyright notice has been invented. Retain this notice, the source model card, the base MIT license, and any additional upstream notices when redistributing.

## Changes

Converted the original Safetensors weights to ONNX opset 17. Applied ONNX Runtime dynamic INT8 quantization to MatMul and Gather weights, per-channel disabled. Model labels, tokenizer, and classification architecture were not changed. Conversion scripts, hashes and environment versions are provided alongside this bundle.

## Scope and limitations

Original authors designate this detector for research and do not recommend production use. Its training covers early ChatGPT outputs in French and English, and its results do not establish reliability for contemporary web content or newer language models. Scores describe this classifier's output; they are not calibrated probabilities of authorship, nor percentages of words actually generated by AI.

Use only as an experimental review signal. Short, edited, translated, formulaic, or out-of-domain text can receive incorrect results. Never treat this result alone as proof that a person used AI. The two included fixtures test that conversion and inference work; they are not an accuracy benchmark.