5 ms·
Wavecoder – a CodeLLM with 6.7B params scoring just behind GPT4
- lxe 3y agoA 6.7B model that's as good as GPT-4 is mostly due to overfitting in such a way that favors certain benchmarks.
- great_psy 3y agoThis a pretty strong claim with zero data to back it up
- eightysixfour 3y agoEvery small model that has outperformed GPT-4 has proven to be an overfit, so I would say it is the obvious claim, and any claim opposite that is what we should be skeptical of.
- anon373839 3y agoWith the exception of task specialization. Fine-tuning a small model such as Mistral 7B on a specific set of tasks can outperform using GPT-4 on those tasks, and with cheaper and faster inference.
- eightysixfour 3y agoNot on the leaderboards mentioned here. That’s my point, you can overfit for specific tasks, you can’t beat them on multi-task leaderboards without training on the test data.
- lxe 3y agoWhile I lack specific data, my intuition is based on observed trends in AI model development. I believe some other models that claimed such numbers excelled in benchmarks but fell short in real-world applications. Further research can validate this claim, and I welcome a balanced discussion.
- travisporter 3y agoIt does seem incredible that chatgpt has so much expertise in literally everything. Does this mean you can beat chatgpt by creating smaller "experts" and directing questions to each?
- great_psy 3y agoSee mixture of experts. It’s likely what chatGPT does in the backend.
- earleybird 3y agoIn their paper they say "To prevent overfitting, we use Low-Rank Adaption (LoRA) [35] for fine-tuning . . ." I'm way out of my league here so I have no opinion on whether or not that actually addresses overfitting. (that quote probably doesn't capture their intention - just a pointer into the paper)
- eightysixfour 3y agoThat’s to prevent overfitting on their dataset, it is not to prevent overfitting on the test data, which is likely in their dataset. You basically cannot beat GPT-4 on broad reasoning tasks, which the tests are designed to cover, without having some of the tests leaking into training dataset. There simply aren’t enough parameters and isn’t enough training to make that possible.