One argument is that it can contaminate training data from output of itself or other models.
We have already documented evidence of the effect of this. In the GPT-4 technical report [1], they reported contamination of humaneval data in the training data.
They did measure against a "non-contaminated" training set but no idea if that can still be trusted.
Why would this matter? We can have seemingly strong benchmarks for containments but measures poorly against new and quarantined information. Classic over fitting.
Another argument is that data being put out there could very much be wrong and the amounts of it amplified by other models. Take a look at this sample of demonstration data for codealpaca [2]. Not only is its output wrong but bad practices like,making up a random computation without it having access to a place to run a calculation, teaches the model these type of responses are ok.
{ "instruction": "What would be the output of the following JavaScript snippet?", "input": "let area = 6 * 5;\nlet radius = area / 3.14;", "output": "The output of the JavaScript snippet is the radius, which is 1.91." }
My guess is that it's because they've already done RLHF on top of the standard next token prediction. In other words, they can't cheaply fine tune ChatGPT without undoing the RLHF objective by training on next token prediction with post-2021 data, and then retraining with RLHF to make sure it still gives good human-like output.
I mention the "undoing RLHF" since it's not uncommon for fine-tuned models to increase in error in the original training objective after being fine-tuned with a different one. I think people saw this happen in BERT.
