Paste text into Detect · Lite for a four-signal heuristic that runs on this page, or open Two-level Detection Lab to walk the open detection pipeline. The lab splits sentences for real and then shows the published response contract. It does not load production weights and it does not score your paste.
Two tiers. This Space is the free CPU layer. The open method is in detection-model-pipeline. The production detector — paste, upload, and sentence-level review — is at lynote.ai/ai-detector.
Open source: github.com/lynote-ai/ai-text-detector
sentence_splitter.py, the fitted temperature T = 1.1842, confidence bands (reject < 0.33 ≤ low < 0.6 ≤ medium < 0.8 ≤ high), and the {code, msg, data} demo response at version V2-hier-1. The banner on that tab says the probabilities are illustrative.The loss sketch L = L_doc + α · L_sent is why a mixed document can be trained: sentence probabilities stay differentiable and feed the document head. This page never runs that forward pass.
The app processes text in memory and does not intentionally store submitted drafts. Hugging Face infrastructure and platform logs remain subject to Hugging Face policies.
MIT. See the linked repositories for their source and license notices.
Paste text into Detect · Lite for a four-signal heuristic that runs on this page, or open Two-level Detection Lab to walk the open detection pipeline. The lab splits sentences for real and then shows the published response contract. It does not load production weights and it does not score your paste.
Two tiers. This Space is the free CPU layer. The open method is in detection-model-pipeline. The production detector — paste, upload, and sentence-level review — is at lynote.ai/ai-detector.
Open source: github.com/lynote-ai/ai-text-detector
sentence_splitter.py, the fitted temperature T = 1.1842, confidence bands (reject < 0.33 ≤ low < 0.6 ≤ medium < 0.8 ≤ high), and the {code, msg, data} demo response at version V2-hier-1. The banner on that tab says the probabilities are illustrative.The loss sketch L = L_doc + α · L_sent is why a mixed document can be trained: sentence probabilities stay differentiable and feed the document head. This page never runs that forward pass.
The app processes text in memory and does not intentionally store submitted drafts. Hugging Face infrastructure and platform logs remain subject to Hugging Face policies.
MIT. See the linked repositories for their source and license notices.