Spaces:
Sleeping
Sleeping
Delete README.md
Browse files
README.md
DELETED
|
@@ -1,67 +0,0 @@
|
|
| 1 |
-
---
|
| 2 |
-
title: groundlens — Hallucination Detection Demo
|
| 3 |
-
emoji: 📐
|
| 4 |
-
colorFrom: yellow
|
| 5 |
-
colorTo: red
|
| 6 |
-
sdk: gradio
|
| 7 |
-
sdk_version: 5.33.0
|
| 8 |
-
app_file: app.py
|
| 9 |
-
pinned: true
|
| 10 |
-
license: mit
|
| 11 |
-
tags:
|
| 12 |
-
- hallucination-detection
|
| 13 |
-
- llm-evaluation
|
| 14 |
-
- rag
|
| 15 |
-
- grounding
|
| 16 |
-
- nlp
|
| 17 |
-
- groundlens
|
| 18 |
-
- embedding-geometry
|
| 19 |
-
short_description: Geometric LLM hallucination detection. No second LLM.
|
| 20 |
-
---
|
| 21 |
-
|
| 22 |
-
[](https://pypi.org/project/groundlens/)
|
| 23 |
-
[](https://github.com/groundlens-dev/groundlens)
|
| 24 |
-
|
| 25 |
-
# groundlens — Hallucination Detection Demo
|
| 26 |
-
|
| 27 |
-
Detects LLM hallucinations using embedding geometry. No second LLM. Deterministic. Auditable.
|
| 28 |
-
Benchmarked against [Vectara HHEM-2.1-Open](https://huggingface.co/vectara/hallucination_evaluation_model).
|
| 29 |
-
|
| 30 |
-
## Methods compared
|
| 31 |
-
|
| 32 |
-
**groundlens SGI** (with context): ratio of Euclidean distances on the embedding space —
|
| 33 |
-
`dist(response, question) / dist(response, context)`. No model inference for
|
| 34 |
-
the evaluation. One embedding call, one division.
|
| 35 |
-
|
| 36 |
-
**groundlens DGI** (without context): cosine similarity between the response
|
| 37 |
-
displacement vector and the mean displacement of verified grounded pairs.
|
| 38 |
-
|
| 39 |
-
**HHEM-2.1-Open** (Vectara): fine-tuned flan-T5 classifier. Full model
|
| 40 |
-
inference per evaluation call.
|
| 41 |
-
|
| 42 |
-
## When they disagree
|
| 43 |
-
|
| 44 |
-
Disagreement surfaces **Type III hallucinations** — factual errors within
|
| 45 |
-
a correct semantic frame. Embedding geometry cannot detect these: the
|
| 46 |
-
response occupies the geometrically correct region of the space despite
|
| 47 |
-
being factually wrong. HHEM's classifier may catch some of these cases.
|
| 48 |
-
The two methods are orthogonal signals, not competing alternatives.
|
| 49 |
-
|
| 50 |
-
## Install the library
|
| 51 |
-
|
| 52 |
-
```bash
|
| 53 |
-
pip install groundlens
|
| 54 |
-
```
|
| 55 |
-
|
| 56 |
-
## Links
|
| 57 |
-
|
| 58 |
-
- [GitHub](https://github.com/groundlens-dev/groundlens)
|
| 59 |
-
- [Documentation](https://docs.groundlens.dev)
|
| 60 |
-
- [PyPI](https://pypi.org/project/groundlens/)
|
| 61 |
-
- [Website](https://groundlens.dev)
|
| 62 |
-
|
| 63 |
-
## Research
|
| 64 |
-
|
| 65 |
-
- [Semantic Grounding Index — arXiv:2512.13771](https://arxiv.org/abs/2512.13771)
|
| 66 |
-
- [Geometric Taxonomy of Hallucinations — arXiv:2602.13224v3](https://arxiv.org/pdf/2602.13224v3)
|
| 67 |
-
- [Rotational Dynamics of Factual Constraint Processing — arXiv:2603.13259](https://arxiv.org/abs/2603.13259)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|