When Bad Data Leads to Good Models
arxiv.org
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.

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.

    • Sculptus Poe@lemmy.worldEnglish
      2·
      2 days ago

      Well, I do wish they would promote the actual use and limitations of AI and stop making up crap and overselling the use cases. I use ChatGPT at work all the time as a start for research, but if I took any of it as being reliable info to run with I would be in grave trouble. It is a great tool that has saved me much time because I know how far to trust it and how to use it. The progress is very impressive as I’ve been using AI art services for years, and the difference between the random blobs from back then and the great stuff it can generate now is pretty stark. Same thing with the LLMs. I’ve been using ChatGPT since it showed up and it has improved greatly since then. Before all this I talked to people who were using AI training on various picture recognition projects where getting data from other sensors was not practical. … Overall AI is pretty exciting, but the non-stop hype and hate headlines is doing nobody any favors.