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.
I’m cool with it. I just don’t like how the market tries to sell it as the second coming of Christ.
This is the same market that tried to add blockchain to everything when that first became well-known.
Some of the biggest forces in the market are extraordinarily stupid people trying to ride every buzzword that comes along.
I think the biggest forces sell the fantasy to smaller forces. This way they can capitalize on the smaller forces believing the hype.
“Don’t believe that marketing department“ is one of those things everybody needs to learn at some point in their life.
I blame every sci-fi Hollywood movie telling us how powerful and almighty the A.I is. How it’s going to be the magic pill that entirely destroys or saves humanity by itself.
Now we have an entire generation believing this crap.
I mean, it still could be. But LLMs are not that AGI we’re expecting.
The difficult question about AGI destroying humanity is deciding whether to be afraid of that option or to cheer it on and LLM enthusiasts are certainly among the people heavily pushing me towards the ‘cheer it on’ option.
You can blame Hollywood for a lot of things, including this, but sci-fi authors have been doing it for longer. That’s where Hollywood took those stories from in the first place.