This is a fine-tuned version of Gemma 3 270M specialized in translating text into emojis. It has been converted to ONNX format to run directly in the browser using Transformers.js.

🎮 Play with it

View Live Demo
(Replace this link with your actual GitHub Pages URL)

💻 Usage in JavaScript (Transformers.js)

You can run this model entirely client-side without a backend!

javascript

import { pipeline } from 'https://cdn.jsdelivr.net/npm/@huggingface/transformers@latest';

// Load the pipeline (uses WebGPU if available, or WASM)
const emojifier = await pipeline('text-generation', 'NathanHannon/emoji_gemma3.270m', {
    dtype: 'q8', // Use the quantized model for speed & low memory
});

// Run inference
// Note: We use manual formatting here to ensure compatibility
const text = "I am so happy to see you!";
const prompt = `<start_of_turn>user\n${text}<end_of_turn>\n<start_of_turn>model\n`;

const output = await emojifier(prompt, {
    max_new_tokens: 20,
    do_sample: true,
    temperature: 0.6,
});

console.log(output[0].generated_text);

🛠 Model Details

Base Model: google/gemma-3-270m-it
Task: Text-to-Emoji Translation
Dataset: kr15t3n/text2emoji
Format: ONNX Weights (Int8 Quantized + FP32)

📁 Files Included

model_quantized.onnx: ~200MB (Recommended for Web/Mobile)
model.onnx: ~1GB (Full FP32 Precision)
tokenizer.json & config.json: Standard tokenizer files

🔧 Training

Finetuned using LoRA (Low-Rank Adaptation) on an RTX 4070.

R: 16
Alpha: 32
Quantization: 4-bit (during training), exported to ONNX Int8.