Thanks so much!

Thanks for the question! I responded in the first post. I'm terrible at responding to the correct threads lol

would absolutely have substitutions for cum and eum because it isn't a language model, rather a traditional deterministic neural net. So in a sense, this small language model behaves /almost/ like a page-level Kraken model, while being an LLM.

This makes the model's output mimic Comma, which is a dataset generated from a deterministic Kraken model. In other words, the LLM's previous training that may or may not have including Matthew would be more representative of the manuscript culture of Comma and the behavior of Kraken which (cont.)

Great point! This is a full-finetune, not just a LoRA adapter. I don't have solid evidence to support this, though it wouldn't be hard to do, but it is pretty clear that a lot of the model's previous training was pushed out of it. We used ~40,000 image/text pairs from the Comma dataset. (cont.)

So for this example, it was the same model. I just changed the temperature. For the other examples, I would train the model with different seeds and/or train the model with different subsets of the data. Each method works, but just slightly differently.

Thanks for sharing!! =) One of the things that you can do to improve this further is train the same model on different subsets of the data or train with different seeds. We're doing that right now and it's showing an improvement of this earlier approach.

Absolutely. It will let you essentially connect named entities (when possible) to linked open data (if available). This means that an individual who appears across a newspaper would have the same identifier.

Thanks! And thanks for the great suggestion! That's something we are testing. The tricky bit is aligning the 3 models. We have 4 models trained on the same data, that may be a good way to do it.

Very cool!! thanks for sharing!

=) me too!

Thanks for your interest! I will very soon!! I'm in the middle of a conference right now, but will try and share a blog post in a week or two.

These models follow the CATMuS guidelines

I should emphasize that these models were taught to not transcribe drop-capitals and marginalia, but they ignore watermarks. Image was from a post by @aaronm.bsky.social

Since each primary source material is extracted from these articles, you can also track how a specific source is used across secondary lit.