yubainu/sibainu-engine
Real-time hallucination detection for LLMs via Geometric Drift Analysis in Hidden States.
What's novel
Real-time hallucination detection for LLMs via Geometric Drift Analysis in Hidden States.
Code Analysis
3 files read · 2 roundsA client-side wrapper that loads a frozen LLM and streams hidden state vectors to an external Hugging Face API for hallucination scoring.
Strengths
The code is functional and demonstrates effective memory management (garbage collection) to run on low-resource hardware. The README accurately describes the system's reliance on an external backend for the core detection logic.
Weaknesses
The project lacks substance as a standalone library because the core algorithm resides in a private/closed API. It has no tests, poor error handling (mostly just breaks loops), and hardcodes specific model IDs and thresholds.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
46
Contributors
1
Files
12
Active weeks
3
Repository
Language
Python
Stars
3
Forks
0
License
NOASSERTION