GLLhJpFfYB's picture
Update README.md
f63841c verified
|
Raw
History Blame Contribute Delete
1.3 kB
---
license: cc-by-4.0
base_model: Sami92/XLM-R-Large-ClaimDetection
tags:
- onnx
- int8
- claim-detection
- check-worthiness
- multilingual
---
# XLM-R-Large-ClaimDetection — INT8 ONNX
A **portable INT8 ONNX** export of
[`Sami92/XLM-R-Large-ClaimDetection`](https://huggingface.co/Sami92/XLM-R-Large-ClaimDetection)
(multilingual check-worthiness / factual-claim detection), for offline CPU inference via
ONNX Runtime.
- **Source model / license:** `Sami92/XLM-R-Large-ClaimDetection`, **CC-BY-4.0** — attribution
required and retained (see the source repo). This is a pure format conversion (fp32 ONNX export
via `optimum`, then portable INT8 dynamic quantisation via
`onnxruntime.quantization.quantize_dynamic`, QInt8), so the CC-BY-4.0 license is **inherited**.
- **`config.json` is byte-identical to the source**, so the label mapping is preserved
(`id2label`: index 1 = `"factual"` = the CLAIM class).
- **Quantisation:** portable QInt8 dynamic — runs through ONNX Runtime's `CPUExecutionProvider`
on any architecture (x86-64 **and** arm64). Deliberately **not** the x86-only `avx512_vnni`
target.
- **Built by** the Buddy framework's `dev/export_gatekeeper_onnx.py` (S3.71), validated to
discriminate an obvious claim from obvious social glue before upload.