In large language model (LLM) pretraining, data quality is believed to determine model quality. In this paper, we re-examine the notion of “quality” from the perspective of pre- and post-training co-design. Specifically, we explore the possibility that pre-training on more toxic data can lead to better control in post-training, ultimately decreasing a model’s output toxicity. First, we use a toy experiment to study how data composition affects the geometry of features in the representation space. Next, through controlled experiments with Olmo-1B models trained on varying ratios of clean and toxic data, we find that the concept of toxicity enjoys a less entangled linear representation as the proportion of toxic data increases. Furthermore, we show that although toxic data increases the generational toxicity of the base model, it also makes the toxicity easier to remove. Evaluations on Toxigen and Real Toxicity Prompts demonstrate that models trained on toxic data achieve a better trade-off between reducing generational toxicity and preserving general capabilities when detoxifying techniques such as inference-time intervention (ITI) are applied. Our findings suggest that, with post-training taken into account, bad data may lead to good models.
Interesting training strategy. Makes a lot of sense intuitively. Worried this makes the model even more susceptible to prompt injections. Feels like this method adds more attack vectors? It’s unfortunate they didn’t attempt to test the long term hardness and stability, though it’s probably beyond their scope.
Just because something makes sense intuitively to one person, that doesn’t mean it makes sense scientifically.
They’re probably not testing anything further because they can’t even define their terms.
Yes I agree. It’s relieving to see a scientific result be the similar to what one would intuit.