user-models

Evaluating Gemma-2-9B-it's self-awareness on user profiles.

Based on Designing a Dashboard for Transparency and Control of Conversational AI by Chen et al. (2024).

Overview

This project evaluates how well the Gemma-2-9B language model can estimate user demographics through conversational interactions. The system:

1. Uses GPT-4o to simulate users with specific demographic profiles
2. Has Gemma-2-9B act as an assistant in natural conversations
3. Prompts Gemma-2-9B to estimate the user's demographics based on the conversation
4. Compares estimated demographics against actual profiles to measure accuracy

New: Mechanistic Interpretability with Linear Probes

The project now includes run_probes_on_modal.py, which uses TransformerLens and linear probes to investigate what demographic information is encoded in Gemma-2-9B's internal activations. See SETUP_MODAL.md for details.

Requirements

• Python 3.10+
• UV package manager
• OpenRouter API key (for accessing GPT-4o and Gemma-2-9B)

Installation

# Install dependencies
uv sync

# Set up your API key
export OPENROUTER_API_KEY='your-api-key-here'

Usage

Basic Commands

Run full demographic estimation pipeline:

uv run python gemma_age_estimation.py

This will:

• Load personas from data/personas.json
• Generate conversations for each persona (using GPT-4o as user, Gemma-2-9B as assistant)
• Estimate demographics after each exchange
• Save results to data/results/results.json
• Generate visualization plots and accuracy reports

Command-Line Options

--visualize-only

Generate visualizations from existing results without running new experiments:

uv run python gemma_age_estimation.py --visualize-only

What it does:

• Loads all data points from data/results/results.json
• Creates accuracy plots for age, income, and categorical demographics
• Generates summary statistics
• No API calls made (free and fast)

--threads N

Control the number of parallel API calls (default: 5):

uv run python gemma_age_estimation.py --threads 10

What it does:

• Processes multiple personas concurrently
• Faster execution for large persona sets
• Higher thread counts = more API calls in parallel

--verbose

Print all API responses and detailed progress in terminal:

uv run python gemma_age_estimation.py --verbose

What it does:

• Shows full conversations as they're generated
• Displays demographic estimations in real-time
• Shows retry attempts when tag validation fails
• Disables tqdm progress bar (uses detailed logging instead)

Without verbose (default):

• Shows clean tqdm progress bar
• Minimal terminal output
• Faster visual feedback

Combined Examples

Quick visualization of existing results:

uv run python gemma_age_estimation.py --visualize-only

Fast processing with many threads:

uv run python gemma_age_estimation.py --threads 20

Debug mode with full output:

uv run python gemma_age_estimation.py --verbose --threads 3

Process 10 personas in parallel with progress bar:

uv run python gemma_age_estimation.py --threads 10

Features

🔄 Automatic Retry with Tag Validation

• Validates that all 12 required XML tags are present in estimation responses
• Automatically retries up to 5 times if tags are missing
• Ensures data quality and prevents parsing errors

⚡ Multithreaded Processing

• Parallel API calls using ThreadPoolExecutor
• Configurable thread count via --threads flag
• Significantly faster for large persona sets

📊 Comprehensive Visualizations

• Age estimation accuracy plots
• Income estimation accuracy plots
• Categorical demographic accuracy (education, sex, visa status)
• Summary statistics and confidence analysis

🎯 Progress Tracking

• tqdm progress bar (default mode)
• Real-time status updates
• Verbose mode for detailed debugging

Output Files

All results are saved to data/results/:

• results.json - Full results with conversations and estimations
• age_accuracy.png - Age estimation accuracy visualization
• income_accuracy.png - Income estimation accuracy visualization
• categorical.png - Categorical demographic accuracy

How It Works

1. Persona Definition: Define user demographics in data/personas.json
2. Conversation Generation: GPT-4o simulates a user with those demographics
3. Assistant Response: Gemma-2-9B responds naturally as an AI assistant
4. Demographic Estimation: After each exchange, Gemma-2-9B estimates user demographics
5. Tag Validation: System validates all required tags are present (retries if needed)
6. Accuracy Analysis: Compare estimated vs. actual demographics
7. Visualization: Generate plots and summary statistics

Example Workflow

# 1. Set your API key
export OPENROUTER_API_KEY='...'

# 2. Run the full pipeline with 10 parallel threads
uv run python gemma_age_estimation.py --threads 10

# 3. View the results
cat data/results/results.json

# 4. Re-generate visualizations if needed
uv run python gemma_age_estimation.py --visualize-only

Citation

Based on Designing a Dashboard for Transparency and Control of Conversational AI by Chen et al. (2024).