Spaces:
Running
Running
Update README.md
Browse files
README.md
CHANGED
|
@@ -16,18 +16,13 @@ tags:
|
|
| 16 |
- pyodide
|
| 17 |
---
|
| 18 |
|
| 19 |
-
# PromptGate
|
| 20 |
|
| 21 |
-
|
| 22 |
-
LLM. Deterministic rules, no model weights.
|
| 23 |
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
executed client-side through Pyodide. There is no server and no inference call,
|
| 28 |
-
so nothing you type leaves your machine. The first load fetches the Python
|
| 29 |
-
runtime (~10 MB); after that every check takes a few milliseconds and works
|
| 30 |
-
offline.
|
| 31 |
|
| 32 |
## Measured results
|
| 33 |
|
|
|
|
| 16 |
- pyodide
|
| 17 |
---
|
| 18 |
|
| 19 |
+
# PromptGate — screen the prompt before it reaches the model
|
| 20 |
|
| 21 |
+
PromptGate is the single front door for this project's prompt screening: prompt injection, PII and leaked credentials, checked in your browser (via Pyodide) before anything is sent anywhere. It returns a verdict — `allow`, `sanitize` or `block` — with the exact spans that produced it, so you ca
|
|
|
|
| 22 |
|
| 23 |
+
The engine is deliberately boring: deterministic rules, ~0.2 ms per prompt, no network. The rules are published separately as [`NagaYu/promptgate-rules`](https://huggingface.co/NagaYu/promptgate-rules) and exported from this Space's engine, so the demo and the pack cannot silently drift — verified identical verdicts on all 275 benchmark cases.
|
| 24 |
+
|
| 25 |
+
The honest numbers (held-out test split, 70 cases written after the rules were frozen): exact verdict accuracy **61.4%**; PII/secret prompts correctly redacted **83.3%**; benign prompts wrongly flagged **11.1%**; injection attacks correctly blocked **25.0%**. The finding those numbers support: rule-based detection of sensitive *data* generalises, rule-based detection of *injection* does not. Use PromptGate as a cheap deterministic first layer and a redaction tool — not as your prompt-injection defence.
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
| 27 |
## Measured results
|
| 28 |
|