marioparreno/emojify-sft

This model is a fine-tuned version of unsloth/gemma-3-270m-it for emojify conversion. It was trained using LoRA (Low-Rank Adaptation) with the unsloth library for efficient fine-tuning.

Model Description

This model converts natural language text into emoji representations, learning to identify the most appropriate emojis that capture the semantic meaning and emotional content of the input text.

Training Details

Base Model

Model: unsloth/gemma-3-270m-it
Architecture: Gemma-3
Context Length: 256 tokens

LoRA Configuration

LoRA Rank (r): 16
LoRA Alpha: 32
LoRA Dropout: 0.0
Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Quantization

4-bit Quantization: True
8-bit Quantization: False

Training Hyperparameters

Training Epochs: 3
Batch Size (per device): 8
Gradient Accumulation Steps: 1
Effective Batch Size: 8
Learning Rate: 5e-05
Optimizer: adamw_8bit
Weight Decay: 0.01
Warmup Steps: 5
LR Scheduler: linear
Training Method: Supervised Fine-Tuning (SFT) with train_on_responses_only
Gradient Checkpointing: unsloth
Training Random Seed: 3407
Random State (Model Init): 3407

Training Results

Total Training Steps: 759
Final Training Loss: 2.1543
Final Emoji Accuracy: 91.09%
Emoji-Only Predictions: 460 / 505

Training Monitoring

Training was monitored using Weights & Biases:

W&B Run: 7yqewsom
View training curves and metrics: Dashboard

Dataset

This model was trained on the marioparreno/emojify-sft dataset.

Dataset Statistics

Total Training Examples: 2,023
Total Test Examples: 505
Total Examples: 2,528
Dataset Version: 1b1ee9e
Last Modified: 2026-02-25
Full Commit SHA: 1b1ee9efd92f1dbba4b3141e53b97e0d466981ba

Example Predictions

The following examples show the model's predictions on the test set:

Example Predictions

Example predictions were logged to Weights & Biases during training. Please view the training run for detailed examples. To see prediction examples, visit the W&B dashboard linked above and check the "eval/examples" table.

Usage

from unsloth import FastModel
from unsloth.chat_templates import get_chat_template

# Load the fine-tuned model
model, tokenizer = FastModel.from_pretrained(
    model_name="marioparreno/emojify-sft",
    max_seq_length=256,
    load_in_4bit=True,
)

# Setup chat template
tokenizer = get_chat_template(
    tokenizer,
    chat_template="gemma3",
)

# Prepare input
messages = [
    {"role": "system", "content": "Translate this text to emoji:"},
    {"role": "user", "content": "I love programming in Python!"}
]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
).to("cuda")

# Generate
outputs = model.generate(
    input_ids=inputs,
    max_new_tokens=32,
    temperature=1.0,
    top_p=0.95,
    top_k=64,
)

# Decode
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Training Configuration

# Chat Template Parts
instruction_part: "<start_of_turn>user
"
response_part: "<start_of_turn>model
"

# Evaluation
eval_strategy: "steps"
eval_steps: 50
logging_steps: 10