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.

  • scrion@lemmy.worldEnglish
    31·
    3 days ago

    Who says you can’t check their outputs? It’s much faster to e. g. read a generated text than to write everything yourself. Same applies to translations, they’ve been excellent for quite a while now.

    Business communication can be handled effortlessly by AI. Of course you read the result before you send it out, but that takes an order of a magnitude less time than formulating and typing all those meaningless sentences.

    And honestly, that’s a perfect use case for AI. I wouldn’t compose a love letter to my family using AI, but a pamphlet, feature description, sales pitch, any bullshit presentation deck? You bet AI excels at those.

    Same applies to content summaries that help augment search indices. Finding a large number of content candidates (e. g. videos) and have AI summarize the contents of said videos to narrow down the search is helpful and works today.

    I’m not looking for AGI. I’m looking for tools to make my life easier, but in an ethical manner that doesn’t advance the destruction of the planet at an exponential rate, just for some tech bro to jerk it and buy another yacht.

    • DeathsEmbrace@lemmy.worldEnglish
      22·
      3 days ago

      You can make a generic fill in the blanks for all of those like I do and just change the key terminology for each scenario. LLMs are competing with search and replace?