citavyn-gemma4-2b-lora
an adapter for enterprise RAG and compliance
1 min readupdated
This lightweight ~95 MB LoRA adapter (a small add-on trained on top of a base model) is built on google/gemma-4-E2B-it. Trained via Supervised Fine-Tuning for enterprise retrieval-augmented generation (RAG) and data governance, it grounds responses strictly in the provided documents, abstains from answering when data is missing, and adheres to regulatory frameworks like GDPR and the EU AI Act. The adapter is aimed at developers and engineers deploying safe language models in corporate environments.
Key features
- Strict source citation grounding using XML blocks
- Refusal to generate answers when the context lacks reliable facts
- Support for security and regulatory standards such as GDPR and NIST AI RMF
- PII masking and protection against prompt injections
What you can do
- Set up secure enterprise search across internal document repositories
- Integrate the adapter via PEFT and multi-tenant vLLM to handle confidential data
- Automate document checks for regulatory compliance
Example
Run the adapter on top of the base Gemma 4 model to answer questions over internal company documentation with mandatory source citations.
Read more
DetailsBase model: google/gemma-4-E2B-it
Format: LoRA adapter (r=16, ~95 MB)
License: Apache License 2.0










Comments