The effect of fine-tuning on language model toxicity