title: SIRIN
emoji: 🔎
colorFrom: purple
colorTo: pink
sdk: docker
app_port: 8501
pinned: false
tags:
- hallucination-detection
- answerability
- longmemeval
- demo
short_description: Hallucination inspection — replay + live API judges
SIRIN — hallucination inspection, live
Interactive demo for SIRIN: A Unified Toolkit for Detecting Contextual Hallucinations in Retrieval-Augmented and Memory-Grounded LLM Systems.
Paste a context–question–answer triple (or pick a curated case from PsiloQA, RAGTruth, or LongMemEval) and see where the answer leaves the evidence: span-level highlighting, response-level verdicts, claim cards, answerability checks, and side-by-side detector comparison.
How detection runs here. This Space has no GPU. The landing cards show verified recorded results from the paper's Qwen3.5-4B campaign (labelled "Recorded result · verified — detection is not live"). Live detection uses API judges: paste your own OpenAI / OpenRouter / Anthropic key in the sidebar — it stays in your browser session only, never stored, logged, or exported. The default judge model is a free OpenRouter route, so a free OpenRouter key is enough to try span-level judging (each span run makes 3 requests; free keys allow ~50/day). No user data is stored.
Full toolkit: https://github.com/sb-ai-lab/SIRIN