google/embeddinggemma-tuning-lab

🤖 EmbeddingGemma Tuning Lab: Fine-Tuning and Mood Reader

This project provides a set of tools to fine-tune EmbeddingGemma to understand your personal taste in Hacker News titles and then use it to score and rank new articles based on their "vibe."

It includes three main applications:

1. 1.A Gradio App for interactive fine-tuning, evaluation, and real-time "vibe checks."
2. 2.An interactive Command-Line (CLI) App for viewing and scrolling through the scored feed directly in your terminal.
3. 3.A Flask App for a simple, deployable web "mood reader" that displays the live HN feed.

✨ Features

• —Interactive Fine-Tuning: Use a Gradio interface to select your favorite Hacker News titles and fine-tune the google/embeddinggemma-300m model on your preferences.
• —Semantic Search Evaluation: See the immediate impact of your training by comparing semantic search results before and after fine-tuning.
• —Data & Model Management: Easily import additional training data, export the generated dataset, and download the fine-tuned model as a ZIP file.
• —Hacker News Similarity Check: View the live Hacker News feed with each story scored and color-coded based on the current model's understanding of your taste.
• —Similarity Lamp: Input any news title or text to get a real-time similarity score (its "vibe") against your personalized anchor.
• —Interactive CLI: A terminal-based mood reader with color-coded output, scrolling, and live refresh capabilities.
• —Standalone Flask App: A lightweight, read-only web app to continuously display the scored HN feed, perfect for simple deployment.

🔧 How It Works

The core idea is to measure the "vibe" of a news title by calculating the semantic similarity between its embedding and the embedding of a fixed anchor phrase, defined in config.py as `MY_FAVORITE_NEWS`.

1. 1.Embedding: The sentence-transformers library is used to convert news titles and the anchor phrase into high-dimensional vectors (embeddings).
2. 2.Scoring: The cosine similarity (or dot product on normalized embeddings) between a title's embedding and the anchor's embedding is calculated. A higher score means a better "vibe."
3. 3.Fine-Tuning: The Gradio app generates a contrastive learning dataset from your selections.
4. 4.Positive Pairs: (MY_FAVORITE_NEWS, [A title you selected])
5. 5.Negative Pairs: (MY_FAVORITE_NEWS, [A title you did not select])
6. 6.Training: The model is trained using MultipleNegativesRankingLoss, which fine-tunes it to pull the embeddings of your "favorite" titles closer to the anchor phrase and push the others away.

🚀 Getting Started

1. Prerequisites

• —Python 3.12+
• —Git

2. Installation

# Clone the repository
git clone https://huggingface.co/spaces/bebechien/news-vibe-checker
cd news-vibe-checker

# Create and activate a virtual environment (recommended)
python -m venv venv
source venv/bin/activate  # On Windows, use `venv\Scripts\activate`

# Install the required packages
pip install -r requirements.txt

3\. (Optional) Hugging Face Authentication

If you plan to use gated models or push your fine-tuned model to the Hugging Face Hub, you need to authenticate.

# Set your Hugging Face token as an environment variable
export HF_TOKEN="your_hf_token_here"

🖥️ Running the Applications

You can run any of the three applications depending on your needs.

Option A: Interactive Fine-Tuning (Gradio App)

This is the main application for creating and evaluating a personalized model.

▶️ To run:

python app.py

Navigate to the local URL provided (e.g., http://127.0.0.1:7860).

Option B: Interactive Terminal Viewer (CLI App)

This app runs directly in your terminal, allowing you to quickly see and scroll through the scored Hacker News feed.

[image]

▶️ To run:

python cli_mood_reader.py

Interactive Controls:

• —[↑|↓] arrow keys to scroll through the story list.
• —[SPACE] to refresh the feed with the latest stories.
• —[q] to quit the application.

You can also start it with options:

# Specify a different model from Hugging Face
python cli_mood_reader.py --model google/embeddinggemma-300m

# Show 10 stories per screen instead of the default 15
python cli_mood_reader.py --top 10

Option C: Standalone Web Viewer (Flask App)

This app is a simple, read-only web page that fetches and displays the scored HN feed. It's ideal for deploying a finished model.

[image]

▶️ To run:

# (Optional) Specify a model from the Hugging Face Hub
export MOOD_MODEL="bebechien/embedding-gemma-finetuned-hn"

# Run the Flask server
python flask_app.py

Navigate to http://127.0.0.1:5000 to see the results.

⚙️ Configuration

Key parameters can be adjusted in config.py:

• —MODEL_NAME: The base model to use for fine-tuning (e.g., 'google/embeddinggemma-300m').
• —QUERY_ANCHOR: The anchor text used for similarity scoring (e.g., "MY_FAVORITE_NEWS").
• —DEFAULT_MOOD_READER_MODEL: The default model used by the Flask and CLI apps.
• —HN_RSS_URL: The RSS feed URL.
• —CACHE_DURATION_SECONDS: How long to cache the RSS feed data.

📂 File Structure

.
├── app.py                  # Main Gradio application entry point
├── cli_mood_reader.py      # Interactive command-line mood reader
├── cli.png                 # Screenshot for CLI app
├── flask_app.py            # Standalone Flask application for mood reading
├── flask.png               # Screenshot for Flask app
├── src/                    # Source code for the application
│   ├── config.py           # Central configuration for all modules
│   ├── data_fetcher.py     # Fetches and caches the Hacker News RSS feed
│   ├── hn_mood_reader.py   # Core logic for fetching and scoring
│   ├── model_trainer.py    # Handles model loading and fine-tuning
│   ├── session_manager.py  # Manages user sessions and application state
│   ├── ui.py               # Defines the Gradio user interface
│   └── vibe_logic.py       # Calculates similarity scores and "vibe" status
├── requirements.txt        # Python package dependencies
├── example_training_dataset.csv # Example dataset for training
├── README.md               # This file
├── artifacts/              # Stores session-specific fine-tuned models and datasets (generated)
└── templates/              # HTML templates for the Flask app
    ├── index.html
    └── error.html