# MGSACore: LLM Interpretability, Attention Salience & Prompt Optimization Framework for OSDR Pipelines **URL:** \ **Category:** AI/ML AWG Topics\ **Tags:** project\ **Created:** [August 20, 2026, 1:28am UTC](https://awg.osdr.space/t/mgsacore-llm-interpretability-attention-salience-prompt-optimization-framework-for-osdr-pipelines/4488 "2026-08-20T01:28:05Z")\ **Posts on this page:** 1\ **Page:** 1
**Author:** ![VanessaMontiel](https://sea1.discourse-cdn.com/flex001/user_avatar/awg.osdr.space/vanessamontiel/32/3858_2.png) [@VanessaMontiel](https://awg.osdr.space/u/VanessaMontiel)\ **Post date:** [August 20, 2026, 1:28am UTC](https://awg.osdr.space/t/mgsacore-llm-interpretability-attention-salience-prompt-optimization-framework-for-osdr-pipelines/4488/1 "2026-08-20T01:28:05Z")
[Vanguardia MGSA - MGSACore Framework](https://colab.research.google.com/github/google-gemini/gemma-cookbook/blob/main/responsible/lit_gemma.ipynb) [MGSACore](https://github.com/Montiel06/MGSACore) Hi [@AIMLawg](https://awg.osdr.space/groups/aimlawg) team, At Vanguardia MGSA, we are developing MGSACore, an open-architecture framework built on top of LIT (Learning Interpretability Tool) and open foundation models (Gemma). Our focus is twofold: Explainability & Safety (MGSACore Shield): Visualizing attention salience token-by-token to prevent hallucinations, detect bias, and ensure research-grade data extraction from scientific literature. Efficiency & Latency (MGSACore Optima): Pruning redundant context in system prompts and RAG retrieval chunks to optimize inference compute. > Attached is our technical One-Pager: > > [One\_Pager\_MGSACore\_EN.pdf](https://awg.osdr.space/uploads/short-url/w0chmMjBQfNewoHWDH5NecUYiMX.pdf) (49.1 KB) We would love to collaborate on active OSDR AI/ML benchmarks or integrate this interpretability workflow into ongoing data projects. Looking forward to discussing this in the next monthly meeting! Best regards, Vanessa Montiel Ruiz | Vanguardia MGSA