omg it almost works!

I stripped out the 5 least used layers. and then I used SFT over 4 epochs and a high learning rate.... and its almost good!

My goal is to make a new Velvet eclipse with these "less used" paramets stripped out. reducing the size significantly to allow for a higher inference speed, and more room for context.

NOTES

        per_device_train_batch_size = 10,
        gradient_accumulation_steps = 4,
        num_train_epochs = 4, # Set this for 1 full training run.
        learning_rate = 5e-4, # Reduce to 2e-5 for long training runs

Uploaded finetuned model

• Developed by: SuperbEmphasis
• License: apache-2.0
• Finetuned from model : SuperbEmphasis/The-Omega-Directive-12B-EVISCERATED

This mistral model was trained 2x faster with Unsloth and Huggingface's TRL library.