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.

  • Naevermix@lemmy.worldEnglish
    91·
    2 days ago

    I envision a Gemini powered bot that cracks captcha and posts “woke” replies on 4chan. If you’re an antivaxxer, antisemite, nazi, racist, sionist, or otherwise, it will debate you. It will not get tired. It will not get mad. It will maintain a sense of decorum indefinitely and it will never ever stop. If some far right extremist decides to do the same, it will have the advantage that academia is left leaning, meaning the model can cite widely recognized studies.

    Dead internet theory and so on, but I’ll gladly completely and utterly destroy the internet if it means the filth dies with it.

    • PushButton@lemmy.worldEnglish
      2·
      2 days ago

      it will have the advantage that academia is left leaning, meaning the model can cite widely recognized studies.

      I was looking for the person saying a particular quote yesterday.

      I asked 3 times the same question and I got 3 different people.

      The funny part us I had the quote wrong.

      Bullshit all the way down.