trained the same way as the original neuralhermes, but using unsloth (2x faster 2x less vram)