The model is based on Google's Gemma 3 270M-IT, a lightweight and efficient language model known for strong instruction-following capabilities in a compact size. The fine-tuning process has specialized this model to take a natural language input (a feeling, situation, or instruction) and output a relevant sequence of emojis.

Base Model: google/gemma-3-270m-it
Parameters: 270 Million
Task: Text Generation (specialized for short-form, emoji-only output)
Architecture: Decoder-only transformer (Gemma 3)

💡 Intended Uses

The primary use case is providing fast, high-quality emoji responses to text inputs.

Text-to-Emoji Translation: Generating a creative and contextually appropriate combination of emojis for a given sentence or emotion.
Resource-Constrained Deployment: Its compact size makes it ideal for fast inference in web applications, on-device (edge) applications, or environments where computational resources are limited.
Research: Experimenting with task-specific fine-tuning on Small Language Models (SLMs).

🛠️ How to Use

The model can be loaded and used with the Hugging Face transformers library. Since this is an instruction-tuned model, it is recommended to use the chat template for prompting.

Installation

bash

pip install transformers accelerate torch

Python Example

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "manuelaschrittwieser/myemoji_generator-gemma-3-270m-it"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16, # Use bfloat16 or float16 if your GPU supports it for faster/lower-memory inference
    device_map="auto"
)

# Define the user prompt for emoji generation
prompt = "Translate the following feeling into exactly 3 emojis: I just finished a very long and difficult project."

# Apply the instruction-tuned chat template
chat = [
    {"role": "user", "content": prompt},
]
formatted_prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)

# Generate output
input_ids = tokenizer(formatted_prompt, return_tensors="pt").to(model.device)

output = model.generate(
    **input_ids,
    max_new_tokens=20, # Keep tokens low since the output is short (emojis)
    do_sample=True,
    temperature=0.7,
    top_p=0.9,
    pad_token_id=tokenizer.eos_token_id # Set pad token to EOS for generation
)

# Decode and print the model's response
response = tokenizer.decode(output[0], skip_special_tokens=True)

# The output might be formatted like a conversation, so we extract the model's part
print(response.split("<start_of_turn>model")[-1].strip()) 
# Example output: "Done! 😮‍💨🏆🥳"

⚠️ Limitations and Ethical Considerations

Scope: This model is highly specialized. It is not suitable for general conversational use, complex reasoning, or factual question-answering.
Bias: The model's associations between text and emojis are a reflection of its training data. It may exhibit biases in its emoji choices or generate culturally inappropriate associations in certain contexts.
Safety: While the base Gemma 3 model has safety mitigations, fine-tuning can alter model behavior. Users should perform their own safety evaluations if deploying the model in a high-risk setting.

📊 Training Details

Fine-tuning Method: This model was fine-tuned using a Parameter-Efficient Fine-Tuning (PEFT) technique (e.g., QLoRA) for maximum efficiency.
Training Data: The model was trained on a custom dataset of (Text, Emoji Sequence) pairs designed to teach the model how to map natural language to emoji output. (Replace this with specific dataset details if you have them, e.g., "Trained on the 5k-sample myemoji_dataset.")
Framework: Hugging Face ecosystem (Transformers, PEFT, TRL).

📄 License

This model inherits its license from the base model. This model is licensed under the Apache 2.0 License.

✍️ Citation

If you use this model in your research, please cite both the original Gemma paper and this model repository.

This Model Citation:

code-snippet

@misc{manuelaschrittwieser_myemoji_generator_2025,
  author = {Schrittwieser, Manuela},
  title = {myemoji\_generator-gemma-3-270m-it},
  year = {2025},
  publisher = {Hugging Face},
  url = {[https://huggingface.co/manuelaschrittwieser/myemoji_generator-gemma-3-270m-it](https://huggingface.co/manuelaschrittwieser/myemoji_generator-gemma-3-270m-it)}
}

Base Model Citation:

code-snippet

@article{gemma2024,
  title={Gemma: Open Models from Google for Responsible AI Development},
  author={Google},
  journal={arXiv:2403.02720},
  year={2024}
}