|
--- |
|
title: Data Protection Team |
|
emoji: 👁 |
|
colorFrom: indigo |
|
colorTo: indigo |
|
sdk: gradio |
|
sdk_version: 4.39.0 |
|
app_file: app.py |
|
pinned: false |
|
license: mit |
|
thumbnail: >- |
|
https://cdn-uploads.huggingface.co/production/uploads/64964a906105da037e718533/aJsEdKX_7CHjvj9ismtAr.jpeg |
|
short_description: I help educators comply with data protection regulations |
|
--- |
|
|
|
<img src="https://i.postimg.cc/3RkhNTY9/ED-COACH-4.jpg"> |
|
|
|
Data protection is paramount when using AI in a school context, but compliance can be difficult when regulations are so complex. |
|
|
|
Fortunately, AI can help! |
|
|
|
This space answers educators' data protection related questions and uses advanced techniques to ensure responses are reliable and actionable. |
|
|
|
<img src="https://i.postimg.cc/0yHMB9W0/Screenshot-2024-07-27-at-17-18-22.png"> |
|
|
|
- It optimizes the user question for retrieval with a light HyDE strategy |
|
- It uses RAG to ground Llama3.1-70B answers in official GDPR, FERPA, and COPPA regulations |
|
- It combines RAG with a long context approach passing the entire documents to Gemini 1.5 Pro |
|
- It uses GPT-4o to synthesize the responses of these two models |
|
- It equips GTP-4o with a Chain-of-Thought system message and instructs it to detects inconsistences / potential hallucinations |
|
|
|
Simple experiments using Claude 3.5 Sonnet as a judge indicate that this approach improves the quality of responses by 5%-10% |
|
compared to a Custom GPT. |
|
|
|
The space is password-protected, but you can |
|
|
|
- Duplicate it and set your own API keys as secret |
|
- Edit the reference documents in app.py if you want to use different / additional regulations |
|
- And even duplicate [this other space](https://huggingface.co/spaces/jeremierostan/Claude_as_a_Judge) if you want to compare its performance |
|
to a Custom GPT such as [this one](https://chatgpt.com/g/g-c9WGlOr9m-data-protection-coach). |
|
|
