Update README.md

README.md

@@ -15,3 +15,57 @@ authors:
 ---
 
 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](https://huggingface.co/GIZ/VULNERABILITY-multilabel-mpnet-multilingual-v2)
+    - Model Name: [https://huggingface.co/GIZ/TARGET-VULNERABILITY-multiclass-mpnet](https://huggingface.co/GIZ/TARGET-VULNERABILITY-multiclass-mpnet)
+    - 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](https://huggingface.co/datasets/GIZ/vulnerability_training_data_full)
+   - 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 dataset 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.
+
+    
+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 via https://huggingface.co/spaces/GIZ/audit_assistant/discussions/new or send us email to [email protected]