Sift

Sift through the noise. Score your feed with EmbeddingGemma, right in the browser.

Inspired by google/embeddinggemma-tuning-lab

Sift is a Chrome extension that runs EmbeddingGemma directly in the browser via Transformers.js + WebGPU:

• EmbeddingGemma-300M (q4) — scores content against your interests using cosine similarity

Current site support (for now): Hacker News, Reddit, and X. More site integrations are planned. On supported sites, Sift dims low-relevance items so the good stuff stands out.

Users can label items (thumbs up/down) to collect training data, export it as CSV, and fine-tune the model with the included Python pipeline.

Demo

sift-demo.mp4

At a Glance

How It Works

1. Category selection — Pick scoring categories like "AI Research" or "Open Source" (25 built-in, custom coming soon)
2. Embedding — Every title/tweet gets embedded by EmbeddingGemma running in your browser (WebGPU/WASM)
3. Scoring — Cosine similarity against category anchor embeddings produces a 0–1 score
4. Category detection — Each item is compared against all active categories; top matches appear as pills in the feed inspector
5. Dimming — Low scores fade out, high scores stay bright. Sensitivity is adjustable
6. Taste profile — After labeling 10+ items, Sift builds a contrastive taste profile showing your top interests ranked by affinity
7. Training — Thumbs up/down on items exports as per-category CSV triplets for fine-tuning

Extension

Chrome MV3 extension in chrome-extension/. Runs entirely client-side — no server, no data leaves your browser.

Supported sites today: Hacker News, Reddit, X (more coming).

Features

• EmbeddingGemma-300M (q4) inference via WebGPU with WASM fallback
• Deterministic score inspector (?) with score band + concise rationale
• Scores HN, Reddit, and X feeds with ambient opacity dimming
• Per-site toggles and sensitivity slider
• 25 built-in categories across tech, world, and lifestyle groups (user-defined categories coming soon)
• Auto-detected category pills — popup hero card and feed inspector show which categories match the current page/item
• Taste profile — contrastive centroid of your positive/negative labels scored against ~100 curated probe phrases; top-5 preview in popup, full-page view with category chips and ranked bars
• Interactive radar chart — SVG spider chart of your taste across all categories; click any axis to drill into that category's sub-topic probes
• Thumbs up/down training labels with per-anchor CSV export
• Label Manager — full-page view to filter, inline-edit, reassign categories, and add labels by URL (auto-fetches title + suggests best category)
• X archive import (like.js, bookmark.js)
• Light/dark mode (follows system)
• Custom model URL for testing local fine-tuned models

Install

1. Download the latest sift-*.zip from Releases
2. Unzip the file
3. Open chrome://extensions, enable Developer Mode
4. Click Load unpacked and select the dist/ folder

Build from source

cd chrome-extension
npm install
npm run build

Then load chrome-extension/dist/ as an unpacked extension in chrome://extensions.

Publishing Fine-tuned Models

Fine-tuned ONNX model files contain only numerical weights, tokenizer vocabulary, and architecture config — no training data, user labels, or personal information. They are safe to publish publicly on HuggingFace Hub, which is the recommended approach for use with the extension.

# Push to HuggingFace Hub after training
python train.py data.csv --push-to-hub your-username/sift-finetuned
# Also valid: --push-to-hub sift-finetuned (auto-uses your username)

Then set the model ID in the extension popup. No authentication is needed for public models.

For private/development models, use the local server instead (python train.py --serve).

Training Pipeline

Fine-tune EmbeddingGemma on your collected labels, export to ONNX, and quantize for browser inference.

Prerequisites

EmbeddingGemma is a gated model — you need to:

1. Accept the license at https://huggingface.co/google/embeddinggemma-300m
2. Create an HF token (read access) at https://huggingface.co/settings/tokens

Colab (recommended if you don't have a GPU)

Open train_colab.ipynb in Google Colab with a T4 GPU runtime. Paste your HF token in the Configuration cell, upload your CSV, run all cells, download the ONNX zip.

Install

uv pip install ".[quantize]"

Usage

# Fine-tune on exported CSV (uses HF_TOKEN env var)
python train.py path/to/sift_training.csv

# Or pass token explicitly
python train.py data.csv --hf-token hf_xxx

# With custom hyperparams
python train.py data.csv --epochs 6 --lr 3e-5

# Convert existing model to ONNX (fp32 + int8 + q4 + q4 no_gather)
python train.py --convert-only path/to/saved_model

# Serve locally for testing in extension
python train.py --serve path/to/onnx_output --port 8000

Set Custom Model URL to http://localhost:8000 in the extension popup to test your fine-tuned model.

Training Data Format

CSV with Anchor,Positive,Negative columns:

Anchor,Positive,Negative
MY_FAVORITE_NEWS,"Show HN: A new static site generator","Ask HN: What's your salary?"
MY_FAVORITE_NEWS,"Rust is eating the world","Bitcoin drops 10%"

Architecture

Content Scripts ──chrome.runtime.sendMessage──▸ Background Service Worker
    (IIFE)       (SCORE_TEXTS, EXPLAIN_SCORE)       (ES module)
                                                       │
Popup ──────────chrome.runtime.sendMessage─────────────┘
    (IIFE)         (GET_STATUS, CATEGORIES_CHANGED)

• Background — Loads EmbeddingGemma via Transformers.js, manages scoring + deterministic inspector rationale, routes messages, stores labels
• Content scripts — Site-specific DOM selectors, MutationObserver for infinite scroll, debounced scoring
• Popup — Settings, category toggles, sensitivity slider, taste profile preview, data export
• Taste page — Full-width taste profile viewer (taste.html), reads cached profile from storage

Project Structure

├── train.py                    # CLI: fine-tune + ONNX export + quantize + serve
├── train_colab.ipynb           # Self-contained Colab notebook (GPU training)
├── pyproject.toml              # Python dependencies
├── src/                        # Python modules (config, trainer, vibe logic)
└── chrome-extension/
    ├── build.mjs               # Dual-build (Vite + IIFE content scripts)
    ├── public/                 # manifest.json, popup HTML/CSS, icons
    └── src/
        ├── background/         # Service worker (model + message routing)
        ├── content/            # HN, Reddit, X content scripts
        │   └── common/         # Shared: batch scorer, styles, widget, labels
        ├── popup/              # Popup UI
        ├── labels/             # Label Manager (full-page label editor)
        ├── taste/              # Full-page taste profile viewer
        ├── agent/              # Agent mode: taste-ranked HN feed (experimental)
        ├── shared/             # Constants, types, taste probes, cache-key utils
        └── storage/            # CSV export, X archive parser

License

Apache-2.0