The Bible is too well-known a text that is too represented in training datasets for this _not_ to be skewed towards poorly reproducing existing translations.
Beyond that,
>there are hallucinations and issues
seems like a deal-killer for a religious text. Yes, all translation by humans is an act of interpretation on some level, and so there's lossiness in all translation – but the difference between a human carefully weighing their reasoning for a particular choice of rendering vs. an LLM that is basically weighted dice that might land totally wrong is a categorically-different thing, not a question of degrees.
This is definitely one area where the training set for the LLM is liable to be polluted by existing translations and even straight memorized english biblical text.
No one is suggesting you replace your ESV or NKJV with this for your religious study. This is as much a technical project of interest as it is a faith-based one.
In terms of your view of the priors on the Bible, you've described in my experience the process all translations go through. We're all skewed by default toward reproducing (poorly) previous translation through word choice modification.
That is, in many ways, the whole thing. My guess is an iterative approach can actually yield a better approach as words shift meaning socially over time.
Not to be too much of a devil's advocate (ha!), but I kindof think I _want_ it to have biblical translation data in the dataset. So long as it's not simply copying something like the KJV or ESV into the output, then this should be a good thing, right?
Because much of what it produces (especially in the "poetic" mode) does seem to be very much "off the beaten path" for a good number of renditions.
I don't think that the goal would be to have a dataset that is completely free from scholarship on the topic of Biblical translation, but rather to synthesize the rules and principles from the collected body of knowledge and apply it (with steering) to the entire Biblical text.
Beyond that,
>there are hallucinations and issues
seems like a deal-killer for a religious text. Yes, all translation by humans is an act of interpretation on some level, and so there's lossiness in all translation – but the difference between a human carefully weighing their reasoning for a particular choice of rendering vs. an LLM that is basically weighted dice that might land totally wrong is a categorically-different thing, not a question of degrees.