4 ms·
The generated text sounds too good; is it possible that the model overfit the source material (especially since the n-previous-tokens value is infinite, while o
by minimaxir 8y ago
The generated text sounds too good; is it possible that the model overfit the source material (especially since the n-previous-tokens value is infinite, while other approaches like char-rnns/textgenrnn use a fixed window length)? It's something I've encountered many times while working with text generation.
- nmca 8y agoTry googling sentences.
- minimaxir 8y agoHmm, I tested a few sentences and it didn't turn up any exact matches (aside to this article), so maybe I'm wrong. With a temperature of 0.7/1.0, that's enough for sufficiently random text I suppose. (the raw, uncurated generated text using the smaller model is a bit more random: https://raw.githubusercontent.com/openai/gpt-2/master/gpt2-samples.txt https://raw.githubusercontent.com/openai/gpt-2/master/gpt2-s...)
- wuthefwasthat 8y agoThose samples are from the large model (GPT-2)! Regarding memorization vs. generalization, see our paper for more analysis.
- yorwba 8y agoHave you tried to determine which parts of the training data contributed to the model generating a certain output? I wonder whether the model avoids reproducing exact matches of the training data by splicing several similar articles together. For example, the generated text about the Civil War mentions that Thomas Jefferson Randolph [0] was named after his grandfather, the president. But is the wording mostly influenced by articles talking about that specific fact, or does it draw from more general examples of someone being named after their grandfather? [0] https://en.wikipedia.org/wiki/Thomas_Jefferson_Randolph https://en.wikipedia.org/wiki/Thomas_Jefferson_Randolph
- zalo 8y agoDoes it seem like there will be any way to go backwards from the sample to the prompt? From a safety perspective, it would be useful to see what prompt a piece of text might have been generated with...
- amvalo 8y agoThe unicorn story struck me as being structurally very similar to the standard wire stories about new species discoveries. That's probably why it's the best of the bunch.
- DenisM 8y agoIt is uncanny. I would expect a language model to pull up different person names for each time that one was called for. For a person name to be consistently used through several paragraphs it is not enough to rely on word co-occurrence. If I had to produce a text like this I would simply take an existing text and replace randomly chosen words with other similar words (as hinted by amvalo). Similar as in - words that tend to occur in similar contexts. So John->Bob throughout entire text. But that would not be a language model product anymore, and where is fun in that? I should set aside some time to read this paper.
- dannyw 8y agoThat’s because the transformer model used has memory.