analytics-agents-uncertainty/da-code-trajectories-viewer

🔍 DA Code Trajectories Viewer

A Streamlit application for visualizing agent trajectories with token-level confidence highlighting. This viewer loads data from a HuggingFace dataset and provides interactive visualizations of model uncertainty.

Features

• —Token-level Confidence Visualization: Color-coded tokens based on logprobs, entropy, or probability gaps
• —Advanced Uncertainty Metrics: Entropy, probability gaps between top-k tokens
• —Window-based Aggregation Analysis: Sliding window analysis to identify uncertainty hotspots
• —Multiple View Modes:
• —Steps View: Collapsible step-by-step exploration
• —Window Analysis: Identify uncertain regions in responses
• —Full Trajectory: Linear conversation flow with highlighting

Dataset

This viewer loads data from analytics-agents-uncertainty/da-code-evaluation-results HuggingFace dataset.

Local Development

Setup

1. 1.Create a virtual environment:

python -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

1. 1.Install dependencies:

pip install -r requirements.txt

1. 1.Set up your HuggingFace token (for private datasets):

# Create .env file
echo "HF_TOKEN=your_token_here" > .env

Run Locally

streamlit run src/streamlit_app.py

The app will be available at http://localhost:8501

Deployment on HuggingFace Spaces

1. 1.Create a new Space on HuggingFace with Docker SDK
2. 2.Push this repository to the Space
3. 3.Add HF_TOKEN as a secret in Space settings (for private datasets)

The Space will automatically build and deploy using the provided Dockerfile.

Environment Variables

Visualization Legend

• —🟢 Green: High confidence (high logprob / low entropy / high prob gap)
• —🟡 Yellow: Medium confidence
• —🔴 Red: Low confidence (low logprob / high entropy / low prob gap)

Colors are dynamically scaled based on each trajectory's actual metric range.