GOgodelski·Sep 28, 2023
Is this code "open source?"[0] It is under a MIT license, has the training scripts, all the data is highly available, etc. But to the best of my knowledge no one has reproduced their results. These people sure couldn't[1] and I'm not aware of any existing work which did. This honestly is quite common in ML and is quite frustrating as a researcher, especially when you get a round of reviewers who think benchmarks are the only thing that matters (I literally got a work rejected twice with a major complaint being about how my implementation didn't beat [0] despite it beating [1]. My paper wasn't even on architecture... so we weren't even trying to improve the SOTA...).
As a researcher I want to know the HPs and datasets used, but they honestly aren't that important for usage. You're right that to "debug" them one method would be to retrain f
Read the rest
rom scratch. But more likely is doing tuning, reinforcement learning, or using a LoRA. Even the company engineers would look at those routes before they looked at retraining from scratch. Most of the NLP research world is using pretrained models these days (I don't like this tbh, but that's a different discussion all together). Only a handful of companies are actually training models. And I mean companies, I don't mean academics. Academics don't have the resources (unless partnering), and without digressing too much, the benchmarkism is severely limiting the ability for academics to be academics. Models are insanely hard to evaluate, especially after RLHF'd to all hell.
> (And sure, maybe you could try to work around with finetuning, or manually patch the binary weights, but that's similar to how people will patch binaries to fix bugs in proprietary software - yes it's possible, but the point of open source is to make it easier)
The truth is that this is how most ML refinement is happening these days. If you want better refinement we have to have that other discussion.
[0] https://github.com/openai/glow
[1] https://arxiv.org/abs/1901.11137