Finetuning details?
What was trained on top of tess?
I used my custom dataset - it's not on HF.
Have you tried the model?
I only tried it for my custom use case and I'm not sure how it works with other cases people might have.
Nope, just downloading it lol.
Is it general purpose? Or like a task specific finetune? Specifically I was thinking about merging it into a general use 200K model.
It might not be very general, I really don't know, haven't tried.
I mostly uploaded it on HF because I wanted to submit it to HF's leaderboard out of curiosity.
Also, please be aware I only tried it in combination with in-context learning, but it just hit me that it might output total nonsense when it's used without in-context learning.
I'm working on a new version and I'll test this one with cases outside the task I finetuned it for.
I just tried it outside my custom use case and without in context learning and it's total rubbish :) .
Interesting. The scores on the HF leaderboard were OK (which is how I found it), maybe because it doesn't use the prompting syntax?
Afaik the HF leaderboard benchmarks use few shots prompting and I think this is the reason why it has ok scores.
With 0 shot it outputs rubbish, with few shots it ouputs something useful.
I uploaded a new model and submitted it to HF leaderboard (I'm waiting for the results).
It still isn't generic enough in a sense that it shouldn't be used for story telling, for example, but only for reasoning and text comprehension.
https://huggingface.co/Mihaiii/Pallas-0.4
I'm closing this discussion. Feel free to (re)open if needed.