5 ms·

Current LLMs are trained on as much data that can be scraped from the public internet. It’s simply not possible to annotate that much data, even with crowdsourc

by rcme 4y ago

Current LLMs are trained on as much data that can be scraped from the public internet. It’s simply not possible to annotate that much data, even with crowdsourcing. It’s not even a matter of cost. You’d basically need to duplicate the amount of data on the internet. I don’t think you’re appreciating the scale of the data involved in training these models.

CuriouslyC 4y ago
Not necessarily. The bloom model (a GPT competitor and similarly sized) was trained on 1.5T of text, which reduces down to 350B unique tokens. If you took a histogram of those unique tokens, it would have a very long tail with probably 1% or less being well represented. That leaves 350M common tokens to serve as the basis for token tuples being fed into crowdsourcing. There are probably ~2-5B very common token sequences, if you had 5 people view each token sequence and give it a few scores, and that process took ~1-2 minutes (these sequences are short), that leaves a conservative estimate of 50 billion person minutes, or ~34 million person days. If you paid these workers $15/hour, that comes out to $12.5 billion dollars, which is not prohibitively expensive for any big tech company when spread out over several years, particularly when it provides a massive competitive advantage.
- rcme 4y ago
  BLOOM isn’t as good as GPT-3 because it doesn’t use as much training data. LLM quality is still data bound [0]. Further limiting data by requiring annotation is not going to work, at least with the current LLM modeling approach. 0: https://www.alignmentforum.org/posts/6Fpvch8RR29qLEWNH/chinchilla-s-wild-implications https://www.alignmentforum.org/posts/6Fpvch8RR29qLEWNH/chinc...
  - CuriouslyC 4y ago
    As the scale of input data goes up linearly, the scale of commonly observed input patterns goes up logarithmically. If we bumped the scale up an order of magnitude in terms of common input tokens, that still means we could annotate the important part of a 150TB text corpus for 125B worth of human annotation. Given that could break the budget of even large corporations, realistically we'd probably train a model to predict the scores of interest using a fraction of that much human annotation, which would be inferior but still a massive improvement. It is also likely that corporations would team up with indirect competitors to share the cost of annotation and gain an advantage against direct competitors.
    - rcme 4y ago
      How do you figure? Let's say a commonly observed input pattern comprises 1% of training data. For a data set of size N, 0.01 * N examples will contain the pattern. If we increase the size to 2N, 0.01 * 2 * N examples will contain the pattern. Why is the growth logarithmic?
      - CuriouslyC 4y ago
        As the data set size increases, the average frequency of non-trivial most frequent patterns will go down, and the tail will get much larger. Thus if you had a 1% cutoff, the percentage of the data set hitting this cutoff goes down as the data set size goes up. Take a look at pareto distributions with high alpha to understand the statistics of it. Of course, this is only true if new data is distinct from old data. If you just copied your data set 10x and pretended it was a 10x larger data set, it would behave like you expect.
        rcme 4y ago
        Hmm I’m still not convinced. Gather training data can be thought of sampling the underlying distribution of the data. In that sense, you’d expect the proportions of things to converge towards the underlying distribution as you gather more data.
        CuriouslyC 4y ago
        That would be true if we were sampling from the underlying distribution in an unbiased and balanced way from the beginning. Instead data is generated and incorporated one set at a time, and each set is biased. Jargon and terms vary, but the language plumbing is the same - new sets bolster common phrases/idioms and lengthen the tail with specific tokens. Keep in mind though, language isn't a stationary process.
        rcme 4y ago
        Even if each dataset is biased, I’m not still not sure how you derived logarithmic growth from the general notion of bias in data. For instance, assuming the data is biased, perhaps it is biased in the other direction and contains more common patterns compared to the underlying distribution.

by rcme 4y ago

CuriouslyC 4y ago

Not necessarily. The bloom model (a GPT competitor and similarly sized) was trained on 1.5T of text, which reduces down to 350B unique tokens. If you took a histogram of those unique tokens, it would have a very long tail with probably 1% or less being well represented. That leaves 350M common tokens to serve as the basis for token tuples being fed into crowdsourcing. There are probably ~2-5B very common token sequences, if you had 5 people view each token sequence and give it a few scores, and that process took ~1-2 minutes (these sequences are short), that leaves a conservative estimate of 50 billion person minutes, or ~34 million person days. If you paid these workers $15/hour, that comes out to $12.5 billion dollars, which is not prohibitively expensive for any big tech company when spread out over several years, particularly when it provides a massive competitive advantage.

rcme 4y ago
BLOOM isn’t as good as GPT-3 because it doesn’t use as much training data. LLM quality is still data bound [0]. Further limiting data by requiring annotation is not going to work, at least with the current LLM modeling approach. 0: https://www.alignmentforum.org/posts/6Fpvch8RR29qLEWNH/chinchilla-s-wild-implications https://www.alignmentforum.org/posts/6Fpvch8RR29qLEWNH/chinc...
- CuriouslyC 4y ago
  As the scale of input data goes up linearly, the scale of commonly observed input patterns goes up logarithmically. If we bumped the scale up an order of magnitude in terms of common input tokens, that still means we could annotate the important part of a 150TB text corpus for 125B worth of human annotation. Given that could break the budget of even large corporations, realistically we'd probably train a model to predict the scores of interest using a fraction of that much human annotation, which would be inferior but still a massive improvement. It is also likely that corporations would team up with indirect competitors to share the cost of annotation and gain an advantage against direct competitors.
  - rcme 4y ago
    How do you figure? Let's say a commonly observed input pattern comprises 1% of training data. For a data set of size N, 0.01 * N examples will contain the pattern. If we increase the size to 2N, 0.01 * 2 * N examples will contain the pattern. Why is the growth logarithmic?
    - CuriouslyC 4y ago
      As the data set size increases, the average frequency of non-trivial most frequent patterns will go down, and the tail will get much larger. Thus if you had a 1% cutoff, the percentage of the data set hitting this cutoff goes down as the data set size goes up. Take a look at pareto distributions with high alpha to understand the statistics of it. Of course, this is only true if new data is distinct from old data. If you just copied your data set 10x and pretended it was a 10x larger data set, it would behave like you expect.
      - rcme 4y ago
        Hmm I’m still not convinced. Gather training data can be thought of sampling the underlying distribution of the data. In that sense, you’d expect the proportions of things to converge towards the underlying distribution as you gather more data.
        CuriouslyC 4y ago
        That would be true if we were sampling from the underlying distribution in an unbiased and balanced way from the beginning. Instead data is generated and incorporated one set at a time, and each set is biased. Jargon and terms vary, but the language plumbing is the same - new sets bolster common phrases/idioms and lengthen the tail with specific tokens. Keep in mind though, language isn't a stationary process.
        rcme 4y ago
        Even if each dataset is biased, I’m not still not sure how you derived logarithmic growth from the general notion of bias in data. For instance, assuming the data is biased, perhaps it is biased in the other direction and contains more common patterns compared to the underlying distribution.