DeformAR Dashboard

This dashboard is part of our EMNLP 2025 System Demonstration submission.
It provides an interactive interface for exploring model outputs, entity-level statistics, error patterns, and similarity matrices across fine-tuned variants.

Built using Dash (Plotly), it allows dynamic navigation between qualitative and quantitative evaluation views, using cached experiment outputs.

---

🔁 Reproducibility Instructions

To run the dashboard with included demo data and outputs:

Step 1: Download the reproducability/ Folder

Download the full folder from the shared Google Drive link below:

🔗 Download reproducability folder from Google Drive

This folder includes:

• ANERCorp_CamelLab_arabertv02/ — raw corpus files
• conll2003_bert/ — alternate dataset (if applicable)
• ExperimentData/ — extracted outputs, model checkpoints, evaluation results
• analysis-config.yaml — config file already pointing to correct folder paths

---

⚙️ Step 2: Ensure the Config Path is Correct

In main.py, this line defines where the config file is loaded from:

CONFIG_PATH = (Path(__file__).parents[1] / "reproducability" / "analysis-config.yaml").resolve()

✅ If you move analysis-config.yaml somewhere else, make sure to update this path in main.py accordingly.

---

Step 3: Run the Dashboard

cd deformar-dashboard
python main.py

📊 Dashboard Structure

The deformar-dashboard/ directory contains a modular Dash application. Key components include:

Folder / File	Description
main.py	Entry point — runs the dashboard
appSettings.py	Initializes Dash app, sets layout and global callbacks
config/	Configuration logic and YAML parsers (e.g., dashboard-config.yaml)
callbacks/	Callback functions for dynamic tab behavior
layouts/	Tab layouts — includes layout managers and tab definitions
managers/	Tab logic managers (e.g., plotting and tab data processing)
notebooks/	Development and exploratory notebooks (miscellaneous/experimental features)
assets/	Static assets (CSS, JS) for styling the Dash app
cache-directory/	Used by Flask-Caching for variant-specific dashboard state

---

🧠 Component Roles

main.py

• Loads analysis-config.yaml to determine variant paths and config
• Starts the Dash server using appSettings.start_app(...)

CONFIG_PATH = (Path(__file__).parents[1] / "analysis-config.yaml").resolve()

📌 Make sure analysis-config.yaml is placed outside the dashboard folder, at the root of the repo (or adjust this path accordingly).

---

appSettings.py

• Creates the Dash app with Bootstrap styling
• Defines all tabs (load, dataset, decision, instance)
• Dynamically enables tabs once data is loaded
• Registers callbacks using DataManager

managers/ vs layouts/

• Managers handle data processing + plotting logic
  • Includes plotting/ utilities and tab_managers/
• Layouts define tab layout components
  • Each tab layout uses its respective manager to fill in data

DataManager and DashboardData

• Loads and caches fine-tuned model outputs and associated metadata
• Provides access to:
  • Token/entity evaluation results
  • Attention and centroid similarity matrices
  • Dataset loaders
  • Model paths and configurations

---

🚀 Running the Dashboard

cd deformar-dashboard
python main.py

• Make sure you've placed analysis-config.yaml at the expected path
• Ensure experiment output files exist and match the config structure

---

🧪 Development Notes

• The notebooks/ folder contains exploratory analysis and in-development features.
• These may later become part of cross-component visualizations in the dashboard.

---