File size: 3,370 Bytes
5040f0c
7418bc9
 
 
 
5040f0c
f963dad
 
5040f0c
7418bc9
7e63b58
 
 
c8f5d39
5040f0c
 
 
3b1ad79
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dcf17a3
 
3b1ad79
 
 
 
dcf17a3
3b1ad79
 
 
 
 
acc4545
3b1ad79
6a47535
3b1ad79
 
 
 
 
 
 
 
 
 
e44261f
3b1ad79
 
 
 
 
 
 
dcf17a3
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
---
title: Climate Vulnerability Analysis
emoji: 🌡️
colorFrom: green
colorTo: gray
sdk: docker
app_file: app.py
app_port: 8501
pinned: false
short_description: Uncover and summarize vulnerable groups findings
authors:
- user: https://huggingface.co/mtyrrell
- user: https://huggingface.co/leavoigt
- user: https://huggingface.co/TeresaK
---

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

### Technical Documentation of the system in accordance with EU AI Act.
System Name: Climate Vulnerability App

Provider / Supplier: GIZ Data Lab & Data Service Center

As of: July 2025


1. General Description of the System
The Climate Vulnerability App is an AI-powered tool to quickly retrieve and summarize relevant information on marginalised groups from (climate) policy documents,
in order to enable users to get a broad overview of the extent to which different marginalised groups are represented in policies. This tool leverages fine-tuned
transformer models to classify references related to pre-determined marginalised groups and an LLM of choice to summarize the most important information that has
been identified. 



3. Model's Used 

   Text Classification:
    - Model Name: VULNERABILITY-multilabel-mpnet-multilingual-v2 (model will be released soon in public, in urgent cases please reach out to us via e-mail)
    - Model Name: TARGET-VULNERABILITY-multiclass-mpnet (model will be released soon in public, in urgent cases please reach out to us via e-mail)
    - Both models mentioned above are fine-tuned versions of the [all-mpnet-base-v2](https://huggingface.co/sentence-transformers/all-mpnet-base-v2)


   Generative LLM used for summaries: 
    -   Model Name: [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct)



5. Model Training Data:

   - The data used to fine-tune the text classification model can be found here: vulnerability_training_data_full (will be released soon, in urgent cases please reach out to us via e-mail) 
   - The training data has been collected by human annotators that are expert in their fields
   - The data does not contain any known bias, however some classes perform better than others (see model card) and risk of potential bias can never be fully excluded.


8. System Limitations and Non-Purposes
    - The system is designed to provide a quick overview of the most relevant information on marginalised groups in climate policy
    - The system is NOT designed to give an in-depth analysis of the document. Output may always be incomplete or falsly classified and should ALWAYS be reviewed by a human.
    - The system does not make autonomous decisions but just provides information.
    - No personal data of users is being processed. 
    - Results are intended for orientation only – not for legal or political advice.

10. Transparency Towards Users
    -	The user interface clearly indicates the use of a generative AI model and the use of text classification transformer models.

    
12. Monitoring, Feedback, and Incident Reporting
    -	Technical development is carried out by the GIZ Data Service Center.
    -	Please reach out through the contact details provided below, if there are any issues or feedback.


13. Contact: For any questions, please contact us via dataservicecenter@giz.de or datalab@giz.de