said-sinhala-ai-dataset/said-hat
Viewer • Updated • 18k • 72
SAID
community
AI & ML interests
None defined yet.
Recent Activity
View all activity
Organization Card
Sinhala AI-text Identification Dataset
The first benchmark for detecting AI-generated text in Sinhala. Human text predates ChatGPT, and AI text was written natively in Sinhala by GPT-4o, DeepSeek-V3 and Gemini 2.5 Pro.
said-hat
Whole-document detection and generator attribution across news, Wikipedia and social media.
8,994 documents · 4,497 human/AI pairs · 3 domains
said-bound
Sentence-level boundary detection in mixed human/AI Wikipedia passages, with aligned all-human originals.
4,244 passages · 39,358 labelled sentences · 1 held-out generator
Highlights
- Three tasks: detection, generator attribution and authorship-boundary localisation.
- Content-matched pairs: every AI text is derived from a real human text, so topic alone is not a shortcut.
- Leakage-free splits: a human text and its AI counterparts always share a split.
- Unseen-generator test: Gemini 2.5 Pro is held out of said-bound's training split.
- False-alarm testing: aligned all-human originals show how often a tagger flags AI in fully human text.
Quick start
from datasets import load_dataset
hat = load_dataset("said-sinhala-ai-dataset/said-hat")
bound = load_dataset("said-sinhala-ai-dataset/said-bound")Released under CC BY-SA 4.0. See each dataset card for sources, construction details and known limitations.
models 0
None public yet