.arunim.fyiTL;DR: I trained a previous-token-predictor, because it is funny! Maybe it can also serve as a jumping-off point for some adversarial robustness helper, kind of like Eliciting Language Model Behaviors with Investigator Agents.
Play with the model here!
GPT says: What happens when you flip the script—literally. Instead of predicting the next token, what if we tried predicting the previous one? Intuitively, pre-trained models trained with the objective of predicting the next token are learning the correlation between input tokens and what follows, and so might be a few steps of fine-tuning away from predicting previous tokens instead. On the other hand, language often follows a somewhat logical structure, and working backwards is often harder—for example, language is probably designed (evolved?) such that there are always many more questions per-answer than there are answers-per-question.
Goals:
[done] Recall how to train a transformer
[done] Flip things around so we learn previous tokens instead of next tokens.
Profit
Implement a bunch of the architectural improvements that I don’t fully understand yet, and use this as an excuse to understand them.
Compare how much easier / more difficult forward vs. backward is in various languages.
Notes
I did the first two, and found that, for small-ish datasets, e.g. shakespeare, the previous token-predictor does quite well (loss < 3.1), while for larger datasets, e.g. open web text, the previous token predictor both requires more substantial training (shocker), and still only achieves a loss of ~3.3.
See wandb logs.
Future extensions
I’m very curious to see how this works for various languages—i.e. what’s the difference in loss for finetuning on a Chinese equivalent of Openwebtext when predicting next tokens vs predicting previous tokens? There are very likely