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.
This is not surprising if you’ve studied anything on machine learning or even just basic statistics. Consider if you are trying to find out the optimal amount of a thickener to add to a paint formulation to get it to flow the amount you want. If you add it at 5%, then 5.1%, then 5.2%, it will he hard to see how much of the difference between those batches is due to randomness or measurement uncertainty than if you see what it does at 0%, then 25% then 50%. This is a principle called Design of Experiments (DoE) in traditional statistics, and a similar effect happens when you are training machine learning models- datapoints far outside the norm increase the ability of the model to predict within the entire model space (there is some nuance here, because they can become over-represented if care isn’t taken). In this case, 4chan shows the edges of the English language and human psychology, like adding 0% or 50% of the paint additives rather than staying around 5%.
At least that’s my theory. I haven’t read the paper but plan to read it tonight when I have time. At first glance I’m not surprised. When I’ve worked with industrial ML applications, processes that have a lot of problems produce better training data than well controlled processes, and I have read papers on this subject where people have improved performance of their models by introducing (controlled) randomness into their control setpoints to get more training data outside of the tight control regime.