I was hoping it was trained ground up on old texts only. Now we don't really know, when it says something archaic, if it's because it's "pretending" to be old school, or because that is what it truly believes.

Probably not enough old texts for it to learn how the language and the world work. Hence the fine tuning.

Sounds like something right up the domain of synthetic data.

There's a sufficient number of old texts - far more than necessary - if we were willing to converse with it in Latin*! Perhaps we could make a translator LLM to converse with the 17th century LLM...

* I'm not talking about Roman-era Latin, Latin was the scholarly/international language for Europe for a long long time. AFAIK, most European Latin literature is untranslated...

Are you sure? In Andrej Karpathy's intro to LLMs[1], he says pre-training uses about 10TB of text from the web.

It's hard to believe that a comparable amount of text had been created, even across all languages, in the entire history of humanity up to the 17th century.

[1] https://www.youtube.com/watch?v=zjkBMFhNj_g

If data quality is good, one can reach results with much less data:

https://arxiv.org/abs/2305.07759

How many stories/tokens did they actually train on? I can't find it in the paper.

Maybe this will help?

https://huggingface.co/datasets/roneneldan/TinyStories

Also note

https://arxiv.org/abs/2309.05463 (which is larger - obviously size still does contribute to performance)

Yes. I think we may have enough for "full finetuning" and erasing to a large extent the previous knowledge. But that's still very far off for pretraining.

"RomeGPT" is next on my list of Monad successors and to give you a general idea, we have on the order of tens of millions of words in classical Latin (and biggest source will… Augustine). There was a BERT Latin project that was able to collect roughly 500 million words in all with mostly early modern and modern Latin.

In comparison I'm currently part of a project to pretrain a French model and we need… 140 billion words.

It's finetuned on [0] OpenHermes-2-Mistral-7B

From the model description

> OpenHermes was trained on 900,000 entries of primarily GPT-4 generated data, from open datasets across the AI landscape.

https://huggingface.co/teknium/OpenHermes-2-Mistral-7B

Yes I needed that for the conversational/instructional capacities. I've made a lot of tests with base models and it would not listen to instruction very well…

It's an LLM, "pretending" and "truly beleiving" are the same, or rather don't exist.

It's like, when you use prompts like : "You're an helpful assistant", is it believing, pretending, or beleiving to be pretending to be an helpful assistant?

It's as funny as disconcerting to see intent and will attributed to probabiltities. Feels sometimes like we're close to making a religion out of this. History of the human race, i guess (https://www.youtube.com/watch?v=xuCn8ux2gbs for the ref).

If LLMs from an internal world model, there are definitely things they "believe", and things they can pretend to believe.

If LLMs have models (tokens), there are definitely tokens inside said models, and a program can manipulate them.

It's more like when you prompt the model with "be an assistant from the time of Shakespeare", you don't know if it'll imitate actual Shakespearean data in its training, or the sea of modern human imitations or "fan fiction".

There are various ways modern knowledge may have "snuck in" through data contamination. If we really want to know what a Shakespearean chatbot would have looked like, we need to cap the training data.

Maybe in the future, using better mechanistic interpretability we can get around this, but not right now.

An ole English LLM would be a delight but is there enough source material to revive dead languages?

Yes you're perfectly right. I've currently tried to maintain some kind of uneasy balance between good conversational capacities (so that it really is a "chatGPT") and cultural reset, which means it may revert from its 17th persona occasionally. Actually this issue is a good illustration that LLM really are latent space explorer. When you prompt a clearly contemporary concept, the default embedding position will shift back to contemporary associations.

As a prompt engineering trick, I find it helps to use faux archaism (such as "pray tell" as an introductory phrase). This is basically a reinforcement anchor in the 17th century region of embedding space.

I wonder if you could get around this by penalising obviously modern words.

Either that or appending archaic expressions in the prompts (a bit like the prompt extension of Midjourney)