DuckyBlender/brainrot-translator-mlx-8bit

This model can effectively translate Brainrot to English and English to Brainrot, with the correct system prompt. Trained in exactly 1 hour on a 16 GB M1 Pro Macbook (8-core) using MLX. This model actually works, and is probably my first actually useful successful fine-tune. I trained Qwen2.5 instead of Qwen3, for simplicity because I had issues with Qwen3, probably because of the thinking modes.

Information

This model DuckyBlender/brainrot-translator-mlx-8bit was converted to MLX format from mlx-community/Qwen2.5-1.5B-Instruct-8bit using mlx-lm version 0.25.2.

Training

mlx_lm.lora --model "mlx-community/Qwen2.5-3B-8bit" --data data --iters 2500 --max-seq-length 200 --num-layers 16 --batch-size 8 --save-every 25 --wandb diegogpt --train

Peak memory: 6.2GB Trained tokens: around 1.1M Dataset is 11326 lines, so trained for ~16 epochs. Calculation:
Total samples seen = iters × batch size = 11326 × 16 = 181216 Epochs = total samples ÷ dataset size = 181216 ÷ 11326 ≈ 16 I've tested the 200th iteration (around 1.5epoch) and it worked similarly well, but here I gave the 2500th iteration.

Charts

Use with mlx

pip install mlx-lm

import mlx.core as mx
from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

def main():
    print("🧠 Brainrot Translator")
    print("=" * 30)
    
    # Ask user for translation direction
    print("Choose translation direction:")
    print("1. English to Brainrot")
    print("2. Brainrot to English")
    
    while True:
        choice = input("Enter choice (1 or 2): ").strip()
        if choice in ["1", "2"]:
            break
        print("Invalid choice. Please enter 1 or 2.")
    
    # Set system prompt based on choice
    if choice == "1":
        system_prompt = "Translate from English to Brainrot"
        direction = "English → Brainrot"
    else:
        system_prompt = "Translate from Brainrot to English"
        direction = "Brainrot → English"
    
    print(f"\nMode: {direction}")
    print("Loading model...")
    
    # Load the model
    model, tokenizer = load("DuckyBlender/brainrot-translator-mlx-8bit")
    # model, tokenizer = load("mlx-community/Qwen2.5-1.5B-Instruct-8bit", adapter_path="adapters")
    
    print("Model loaded! Enter text to translate (Ctrl+C to exit)")
    print("-" * 50)
    
    try:
        while True:
            # Get user input
            user_input = input(f"\n[{direction.split(' → ')[0]}]: ").strip()
            
            if not user_input:
                continue
            
            # Format the prompt
            messages = [
                {"role": "system", "content": system_prompt},
                {"role": "user", "content": user_input}
            ]
            
            # Apply chat template
            prompt = tokenizer.apply_chat_template(
                messages, 
                tokenize=False, 
                add_generation_prompt=True
            )

            sampler = make_sampler(temp=0.5, top_p=0.95)
            
            print(f"[{direction.split(' → ')[1]}]: ", end="", flush=True)
            
            # Generate response
            response = generate(
                model,
                tokenizer,
                prompt=prompt,
                sampler=sampler,
                max_tokens=128,
                verbose=False
            )
            
            print(response)
            
    except KeyboardInterrupt:
        print("\n\nGoodbye! 👋")

if __name__ == "__main__":
    main()