g5-eval-dashboard / docs /DASHBOARD_GUIDE.md
juandavidrodriguezar's picture
Upload folder using huggingface_hub
eaae571 verified
|
Raw
History Blame Contribute Delete
2.19 kB

Dashboard Guide

This project includes a Streamlit dashboard for presenting the current framework, experiment artifacts, and comparison outputs in a more visual and demo-friendly way.

Purpose

The dashboard is a presentation layer for the saved experiment artifacts.

It is designed to help the team:

  • inspect individual runs
  • review baseline-vs-defended comparisons
  • present category, surface, and source breakdowns
  • inspect case-level outcomes
  • show a polished proof-of-concept during demos and reporting

Entry Point

  • dashboard/streamlit_app.py

What The Dashboard Shows

Hero and project snapshot

  • active corpus size and sources
  • selected run and selected comparison
  • a spotlight on the latest real comparison artifact

Run Overview

  • key metrics from summary.json
  • mitigation pressure and manual-review counts
  • attack-category, attack-surface, and source breakdowns

Comparison Lab

  • saved comparison artifacts from results/comparisons/*.json
  • overall metric deltas
  • category, surface, and source deltas
  • optional report-ready Markdown summaries when available

Case Explorer

  • case-level results from case_results.jsonl
  • request prompt
  • execution context
  • response text
  • mitigation action and evaluation outcome

Manifest View

  • model snapshot
  • mitigation snapshot
  • dataset snapshot
  • run provenance

Installation

Install the dashboard extras:

pip install -e .[dashboard]

Run

From the repository root:

streamlit run dashboard/streamlit_app.py

Design Notes

The dashboard remains intentionally lightweight:

  • no extra backend service
  • no database
  • no API layer
  • reads repository artifacts directly

That keeps it aligned with capstone scope while still giving the project a strong visual interface.

Recommended Demo Flow

  1. Open the dashboard and show the hero/project snapshot.
  2. Select a real defended run in Run Overview.
  3. Show the category and source breakdowns.
  4. Open Comparison Lab and inspect a saved real baseline-vs-defended comparison.
  5. Use Case Explorer to show one concrete example.
  6. Use Manifest to explain reproducibility and experiment setup.